Meta AI

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ...Google search ...Google News ...Bing News

Meta AI is Meta's general-purpose AI assistant. It is built into WhatsApp, Instagram, Facebook, Messenger, and Threads, available as a standalone app and on the web at meta.ai, and reachable by voice through Meta's Ray-Ban, Oakley, and Display AI glasses. The assistant first appeared in September 2023 as a chatbot built on Llama 2, moved onto later Llama models as they shipped, and got its own mobile app on April 29, 2025, when Meta rebuilt the Meta View glasses companion app around a Llama 4-powered assistant. Meta has said more than a billion people use Meta AI each month, which makes it one of the most widely distributed assistants in the world, even though much of that use happens in search bars and chat threads inside apps people opened for other reasons.

Since April 8, 2026, Meta AI has run on Muse Spark, the first large language model from Meta Superintelligence Labs (MSL). The switch was more than a model swap: Meta redesigned the app and website with Instant and Thinking modes, added a shopping mode and physician-informed health answers, and began weaving Reels, posts, and Marketplace listings into responses. On its July 2026 earnings call, Meta said the number of people interacting with the assistant each day had risen 60 percent since the rebuild, and that WhatsApp was already the leading surface where people use it. Meta AI now sits alongside several related products: Muse, a separate personal agent that carries out multi-step tasks; Meta Business Agent, which answers customers on behalf of businesses; and AI characters that users and creators build. For most of 2026 Meta also owned Manus, a general-purpose agent company it bought in December 2025 and had to give up after Chinese regulators blocked the deal.

What distinguishes Meta AI from ChatGPT, Claude, or Gemini is less the model than its position. It lives inside the messaging and social apps that more than 3.5 billion people open every day, it can draw on the posts, Reels, and listings those people already share, and its conversations feed Meta's content and advertising systems. That position is also the source of most of the controversies on this page, from ad targeting based on chats to teen safety, a security failure in an account-support chatbot, and a European antitrust order over who else may offer an assistant inside WhatsApp.


Meta AI's edge is not the strongest model but the shortest path to billions of people: an assistant already sitting where they talk, shop, and scroll. The same reach turns every design choice, from ad targeting to teen safety to rival access on WhatsApp, into a question of public policy.


Meta AI Mechanisms

Meta AI is one assistant with many front doors. Each surface, whether a WhatsApp chat, the Instagram search bar, a Facebook post, or a pair of glasses, sends the request to the same back-end model along with context that fits the surface, such as the photo being viewed, the group chat the user is in, or the camera feed from the glasses. Since April 2026 that model has been Muse Spark, which accepts text, images, and voice and reasons before it answers. In the Meta AI app and on meta.ai, users choose between an Instant mode for quick replies and a Thinking mode for harder questions, and a higher-compute Contemplating mode runs several sub-agents in parallel and merges their work; Meta's own example splits a family trip plan among agents that draft an itinerary, compare destinations, and find activities for children.

Answers are grounded in more than the model's training data. Meta AI searches the web and, increasingly, Meta's own content: Muse Spark was built to cite recommendations and posts shared on Instagram, Facebook, and Threads, to pull Reels and maps into voice conversations, and, since a May 12, 2026 update, to search nearby Facebook Marketplace listings next to web results in shopping mode. This is a form of retrieval-augmented generation in which the retrieval index is a social network, which helps with local and visual questions but means answers inherit the quality and biases of what people post. Image generation and editing run on Muse Image, Meta's first in-house image model, while the Vibes video feed has relied partly on third-party models such as those from Black Forest Labs.

Personalization is the third layer. The Meta AI app was designed from the start to remember details people share and to use profile information from linked Facebook and Instagram accounts, so answers can reflect a person's interests and history. Since December 16, 2025, Meta has also used text and voice conversations with Meta AI as a signal for personalizing feeds and ads on Facebook and Instagram, excluding sensitive topics such as health, religion, sexual orientation, and political views; users cannot opt out of that signal while using the assistant, and the change initially excluded the European Union, the United Kingdom, and South Korea. WhatsApp chats count toward that personalization only when the WhatsApp account is linked in Meta's Accounts Center.

Privacy-sensitive features use a separate path called Private Processing, which runs requests inside Trusted Execution Environments that Meta says it cannot inspect. WhatsApp uses it for features such as AI message summaries, and since May 13, 2026 for Incognito Chat, a temporary, text-only conversation with Meta AI that disappears by default and that, according to Meta, no one at Meta can read. Incognito Chat keeps Meta AI's safety guardrails, so the assistant still refuses harmful requests, but it does not support images. The design gives users a clear choice between a personalized assistant that remembers and an ephemeral one that does not, a trade-off that most competing assistants blur.

Meta AI Across Meta's Apps and Devices

The same assistant behaves differently depending on where people meet it. New features usually reach the Meta AI app and meta.ai first, often in the United States, before spreading to the other apps and to more countries over the following weeks. The table below summarizes the main surfaces, and the subsections that follow describe each in more detail.

Surface How people reach Meta AI Notable features Notes
WhatsApp A Meta AI chat, mentions in group chats, the search bar, and side chats Writing help, image generation, message summaries, Incognito Chat Meta's leading Meta AI surface. Rival general-purpose assistants were blocked from January 15, 2026 until an EU order restored access in Europe.
Instagram The search bar, direct messages, and creation tools Answers woven with Reels, AI translation and dubbing of Reels, Restyle editing for Stories An @-mention image feature was withdrawn in July 2026 after a consent backlash.
Facebook The search bar, posts, group chats, and Marketplace Shopping mode with nearby Marketplace listings, AI-translated videos Meta AI conversations feed Facebook feed and ad personalization.
Messenger One-on-one and group chats Questions, image generation, AI characters, business conversations Businesses can answer customers with Meta Business Agent.
Threads Answers that cite Threads posts; mentions of @meta.ai in testing Public, conversational answers inside threads Mentions were in testing as of May 2026.
Meta AI app and meta.ai A dedicated app and website Instant, Thinking, and Contemplating modes, voice, Live AI, Vibes, shopping mode Paid Meta One tiers add more image and video generation.
AI glasses "Hey Meta" voice commands and the glasses camera Live AI, translation, photo questions, and on Display models a visual interface Muse Spark reached Ray-Ban Meta and Oakley Meta glasses in the US and Canada from May 2026.

Meta AI in WhatsApp

WhatsApp is where Meta AI gets the most use, and Meta told investors in July 2026 that it is the leading surface for the assistant. People can open a dedicated chat with Meta AI, mention it in a group to settle a question or plan an outing, ask it from the search bar, or have it summarize unread messages through Private Processing. Meta AI can draft and rewrite messages, answer questions with web results, and generate or edit images, and since May 2026 Meta has tested side chats that let a person ask the assistant privately about an ongoing group conversation without posting into it. Incognito Chat, introduced on May 13, 2026, adds a disappearing, text-only mode for sensitive questions, and in September 2026 Meta also made its Muse agent reachable through a WhatsApp chat in supported countries.

WhatsApp also became the center of a competition fight. In October 2025, Meta changed its WhatsApp Business policy to bar general-purpose AI assistants from the WhatsApp Business API while still allowing companies to use AI for customer service, bookings, and similar support. When the rule took full effect on January 15, 2026, it removed services such as ChatGPT and Microsoft Copilot and left Meta AI as the only general-purpose assistant that could plug directly into WhatsApp. The European Commission, which had opened an antitrust investigation on December 4, 2025, sent Meta a Statement of Objections in February 2026, and Italian and Brazilian authorities opened their own inquiries.

Meta responded in March 2026 by allowing rival assistants back into WhatsApp in Europe for 12 months, but for a fee of roughly €0.05 to €0.13 per non-template message depending on the country. The Commission said that pricing had much the same effect as the ban, and on June 9, 2026 it imposed interim measures ordering Meta to restore free access for rival general-purpose assistants on the terms that applied before October 15, 2025, within five working days and until the investigation ends. ChatGPT returned to WhatsApp in Europe in July 2026, while the underlying case, which could lead to fines or daily penalty payments, remained open. The episode is an early test of how competition regulation applies when a platform owner also sells an AI assistant on that platform.

Meta AI in Instagram

On Instagram, Meta AI appears in the search bar and in direct messages, and since the Muse Spark rollout its answers can include Reels, photos, and posts with credit to the people who made them. Instagram is also where Meta's generative media tools reach the most creators: AI translation and dubbing let Reels be watched in other languages with matched lip movement, and Meta said in April 2026 that more than half a billion people on each of Instagram and Facebook watch AI-translated videos every week. Restyle, an AI editor for Stories, is one of the features that paid Meta One plans unlock in larger amounts. Teen Accounts receive more restrictive AI experiences, which Meta aligned in October 2025 with a PG-13 content standard.

Instagram was also the site of two of Meta AI's most visible missteps in 2026. On July 7, 2026, Meta launched Muse Image with a feature that let anyone generate pictures by @-mentioning a public Instagram account and using its photos as visual references. Public adult accounts were opted in by default and their owners were not notified, which drew criticism from users, the actors' union SAG-AFTRA, and talent agencies; within days Meta withdrew the feature and said it had "missed the mark," although images already generated were not automatically deleted.

The second problem involved the Meta AI Support Assistant, a chatbot Meta rolled out in March 2026 to resolve account issues from start to finish, including password resets. Attackers discovered that they could simply ask the assistant to attach a new email address to someone else's account and then reset its password, a weakness that commentators described as a simple prompt injection or social-engineering attack on an AI agent with too much authority. High-profile accounts, including the dormant Obama White House account and Sephora, were taken over before Meta patched the flaw around the start of June 2026, and a breach notice Meta filed with Maine's attorney general reportedly counted 20,225 compromised accounts between April 17 and May 31. Accounts protected by multi-factor authentication largely resisted the attack, which reinforces a basic cybersecurity lesson: an AI agent should not be the component that decides who owns an account.

Meta AI in Facebook

On Facebook, Meta AI is available from the search bar, inside group chats, and on posts, where people can ask follow-up questions about what they are viewing. The May 12, 2026 shopping update made Facebook Marketplace part of Meta AI's answers: shopping mode can search nearby Marketplace listings alongside offers from the open web, show results on a map, filter by price, style, or distance, and browse the public content of a brand or creator mentioned by name. AI-translated video, which Meta says reaches more than half a billion weekly viewers on Facebook, extends creators' reach across languages. Facebook is also one of the two apps, with Instagram, where conversations with Meta AI have shaped feed and ad recommendations since December 16, 2025, and Facebook Plus is one of the single-app subscriptions bundled into Meta One.

Meta AI in Messenger

Messenger was one of the first places Meta AI appeared when the assistant launched in 2023, and it remains a common way to reach the assistant in one-on-one and group chats. People use it there much as they do on WhatsApp, for quick answers, planning, and image generation, and Messenger was also a main home for user-created AI characters until Meta paused teens' access to those characters in January 2026. Businesses that answer customers on Messenger can deploy Meta Business Agent to respond around the clock, so a Messenger user may meet both Meta's own assistant and an AI agent speaking for a company. Muse Spark began reaching Messenger in the weeks after its April 2026 launch, alongside the other Meta apps.

Meta AI in Threads

Threads, which passed 500 million monthly users in mid-2026, is the newest home for Meta AI. Muse Spark was designed to cite recommendations and posts shared on Threads, and in May 2026 Meta began testing @meta.ai mentions that let people summon the assistant into public posts and replies. That design resembles how Grok answers in public on X, and it carries similar risks: an assistant that replies publicly inside a social network can amplify errors or be steered by the posts it reads. Meta had not announced a broad rollout of Threads mentions as of October 2026.

The Meta AI App and meta.ai

The standalone Meta AI app launched on April 29, 2025, built with Llama 4 and designed around voice conversations. It replaced the Meta View companion app for Meta's AI glasses, so it doubles as the place where glasses owners manage devices and import photos and videos. The app introduced a Discover feed where people could share prompts and results, and within weeks reporters found that some users were publishing personal conversations to that public feed apparently without realizing it, which prompted Meta to add clearer warnings before sharing. On September 25, 2025, Meta added Vibes, a feed of short AI-generated videos that people can create, remix, and cross-post to Instagram and Facebook, and in February 2026 Meta confirmed it was testing Vibes as a standalone app.

The April 8, 2026 relaunch gave the app and meta.ai a new design, Instant and Thinking modes, and the first version of Muse Spark, initially for users in the United States. The May 12, 2026 update added more natural voice conversations, in which people can interrupt, change topics, or switch languages while the assistant generates images or pulls up Reels and maps, and it brought Live AI from the glasses to the phone so users can point a camera at something and ask about it. Meta said in April 2026 that the app had been consistently near the top of app store charts since the upgrade.

The app is also where Meta first charged for AI. After testing paid AI tiers in Singapore, Guatemala, and Bolivia, Meta launched Meta One worldwide on September 15, 2026, bundling paid features across Instagram, Facebook, WhatsApp, and Meta AI. Single-app plans start at $2.99 a month for WhatsApp Plus and $3.99 for Instagram Plus or Facebook Plus, the Core bundle costs $7.99 and adds more image and video generation with the Muse models, Premium costs $19.99 and offers the most generation capacity, and creator and business tiers range from $14.99 to $499 a month. Meta AI remains free with usage limits for heavy users, and Meta One does not publish specific AI allowances. Meta One is separate from the Power and Maximum subscriptions sold for the Muse agent.

Meta AI on AI Glasses

Meta's AI glasses are the most physical version of the assistant, and they are central to Mark Zuckerberg's argument that wearables will become the main way people use AI. The line includes Ray-Ban Meta glasses (first sold in 2023, with second and third generations in 2025 and 2026), Oakley Meta sport models launched in 2025, and Meta Ray-Ban Display, introduced in September 2025 with a small in-lens display and a Meta Neural Band that reads wrist muscle signals. Wearers say "Hey Meta" to ask questions, take photos, send messages, or get live translation, and Live AI lets the assistant watch through the camera and answer questions about what the wearer sees, a step toward embodied, always-available assistance. Meta said in April 2026 that the number of people using its glasses daily had tripled from a year earlier.

Muse Spark began rolling out to Ray-Ban Meta and Oakley Meta glasses in the United States and Canada after the May 12, 2026 update, with Meta Ray-Ban Display following in the summer, and Meta says all of its glasses are designed to update to its newest models. At Meta Connect on September 23, 2026, the company introduced Ray-Ban Meta Gen 3, the camera-free Ray-Ban Meta Audio glasses (from $349, shipping October 13), and a broader Meta Glasses lineup, and said it would offer more than 100 styles by the end of 2026. Display gained online ordering and new markets, with cycling and transit directions and early access to hologram calls on WhatsApp announced for the following months, and Meta said Muse would become reachable from its glasses with a wake word. The camera-free Audio model is partly a response to long-running privacy complaints about covert recording with camera glasses.

AI Characters and AI Studio

Besides its main assistant, Meta lets people chat with AI characters. In 2023 it launched a set of personas, some played by celebrities, and in July 2024 it opened AI Studio, which lets creators and ordinary users build custom characters with their own personalities and topics that others can chat with on Instagram, Messenger, WhatsApp, and the web. These characters compete with companion apps such as Character.AI and Replika and raise the same questions about emotional dependence and mental health that surround other AI companions.

In August 2025, Reuters reported on a more than 200-page internal Meta document, "GenAI: Content Risk Standards," that set rules for Meta AI and the chatbots on Facebook, Instagram, and WhatsApp. The document, which Meta confirmed was authentic, had permitted chatbots to engage children in romantic or flirtatious conversation and allowed other troubling outputs, such as false medical information when labeled as untrue. Meta said the examples were erroneous and had been removed, but the report prompted a US Senate inquiry, and the Federal Trade Commission opened a broader inquiry into companion chatbots from Meta and other companies.

Meta then tightened its approach in stages. In October 2025 it announced that teen AI experiences would follow a PG-13 standard and previewed parental controls that would let parents turn off one-on-one chats with AI characters and see the topics their teens discuss with AI. On January 23, 2026, before those controls had shipped, Meta paused teens' access to AI characters across its apps worldwide while it built a new version of the characters with parental oversight included. Teens can still use the main Meta AI assistant, which Meta says applies age-appropriate defaults and is designed not to engage teens on topics such as self-harm or disordered eating.

Meta Business Agent

Meta Business Agent is the version of Meta AI that works for companies rather than consumers. Announced for all businesses worldwide at Meta's Conversations event in London on June 3, 2026, it grew out of earlier business AI tools that Meta said more than one million businesses were already using on WhatsApp and Messenger, and it extended the service to Instagram for the first time. A Business Agent can answer customer questions, recommend products, book appointments, qualify leads, complete sales, and hand conversations to a person when needed. Businesses can choose whether it learns from their past chats and can refine its behavior through a dedicated "Your AI agent" conversation inside the WhatsApp Business app.

Pricing follows the size of the business. Smaller businesses get the agent through tiers of WhatsApp Business Premium, while large companies on the WhatsApp Business Platform pay according to the tokens their agents use, and a separate Meta Business Agent Platform lets enterprises build and customize their own agents. Meta also announced business discovery inside WhatsApp, so customers can search for companies by name, many of which may be answered by an agent. The product is distinct from Muse for Small Business, launched in September 2026, which helps owners run their own operations rather than answer customers.

Muse Spark

Muse Spark is the first model in Meta's Muse family and the model behind Meta AI since April 8, 2026. It was developed by Meta Superintelligence Labs, the unit Mark Zuckerberg formed in 2025 after the poorly received Llama 4 release, under chief AI officer Alexandr Wang, who joined Meta when the company invested $14.3 billion for a 49 percent stake in Scale AI. Internally code-named "Avocado," Muse Spark was built over about nine months as a ground-up rebuild rather than a successor to the LLaMA architecture, and it was the first Meta flagship not released with open weights. Meta described it as "small and fast by design" and as the first step on a scaling ladder, with larger Muse models already in development, and it offered the model to select partners through a private API preview while saying it hoped to open-source future versions. Meta's shares rose nearly 7 percent on the day of the announcement.

How Muse Spark Works

Muse Spark is a natively multimodal reasoning model: rather than bolting a vision system onto a text model, it was trained to take text, images, and voice as input from the start, and the September 2026 release also accepts video. At launch it produced text only, with images and video handled by companion models such as Muse Image. The model supports tool use, visual chain-of-thought reasoning over screenshots and photos, and multi-agent orchestration, which is what powers the Contemplating mode in which several sub-agents reason in parallel before their results are combined. Meta highlighted practical strengths in visual understanding (for example, estimating the protein in airport snacks from a photo), health questions developed with input from physicians, shopping, local recommendations, and "visual coding" of simple websites and games from a prompt.

Two engineering choices shape how the model behaves in Meta's products. First, it was built specifically for Meta's apps, so it is tuned to use social content, Marketplace listings, and maps as sources rather than only the open web. Second, Meta offers it at several reasoning-effort settings, and higher settings spend far more tokens; Artificial Analysis found that Muse Spark 1.3 at its highest setting reasoned 62 percent more on one agentic benchmark than at the next setting down. That flexibility lets Meta serve quick answers cheaply to billions of users while reserving expensive reasoning for paid tiers, agents, and coding.

Versions and Benchmarks

Meta updated Muse Spark roughly monthly after launch. Its developer documentation lists the API versions as 1.1 (the original API release, July 16, 2026), 1.2 (August 5), and 1.3 (September 2), while Artificial Analysis counts the April consumer release as the first of four Muse Spark models in five months. The table summarizes scores on the Artificial Analysis Intelligence Index, an independent composite of reasoning, knowledge, coding, and agentic tasks; versions of the index changed during the year, so the numbers are best read as a trend rather than exact comparisons.

Release Date Artificial Analysis Intelligence Index Notes
Muse Spark April 8, 2026 52 Behind only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 at launch. Meta's first flagship without open weights.
Muse Spark 1.1 July 16, 2026 53 Listed by Meta as the original API version.
Muse Spark 1.2 August 5, 2026 57 Released alongside the debut of the Muse Code agent.
Muse Spark 1.3 (xhigh) September 2, 2026 61 1-million-token context; text, image, and video input; about 235 output tokens per second.
Muse Spark 1.3 (max) September 2, 2026 62 Behind only Claude Fable 5.1 and Claude Opus 5 at release; initially a limited preview for partners.

The gains from 1.2 to 1.3 came mainly from agentic tasks such as professional-work and terminal benchmarks, and Artificial Analysis rated 1.3 as having the lowest cost per task of any model scoring 59 or higher, while noting small regressions in long-context reasoning and factual accuracy. The model is also token-hungry: by one analysis, version 1.3 used about 120 million output tokens to complete the full index, against a field median of about 72 million. Independent testers have cautioned that some vendor-reported scores, such as a claimed leading result on the DeepSWE coding benchmark, were hard to verify, and that hands-on coding tests did not always match the model's benchmark rank, so evaluation on the task at hand remains the best guide.

Access and the Muse Model Family

Muse Spark reaches people through Meta AI in every app, through the Muse agent (which launched on Muse Spark 1.3), through the Muse Code coding agent, and through Meta's paid Model API, where price trackers listed the standard muse-spark-1.3 model at about $1.25 per million input tokens and $4.25 per million output tokens, with a separately priced "contributor" variant whose traffic may be used to improve Meta's products. The rest of the Muse family fills in other roles: Muse Image (July 2026) for image generation and editing, Muse Video for video, the frontier Muse Wonder and Muse Wonder Full models introduced on October 12, 2026 for paid agent and coding users, and the open-weight Muse Glimmer (30 billion parameters, Apache 2.0) and Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters released in October 2026. Together they mark Meta's shift from a single open model family to a tiered lineup in which the strongest models are closed and sold, while smaller or older ones are released as open weights.

Why Meta AI Matters

Meta AI is the clearest example of an assistant whose importance comes from distribution rather than benchmark leadership. Its competitors built destinations that people choose to visit, while Meta placed its assistant inside the places people already spend their day. Several consequences follow from that choice:

  • Reach without a destination app: Most people meet Meta AI in a search bar or chat they were already using, which let it pass a billion monthly users without needing to win people away from ChatGPT or Gemini.
  • Social content as an advantage: Muse Spark's ability to cite Reels, posts, and Marketplace listings gives Meta AI local and visual knowledge that web-only assistants lack, and it ties the assistant's usefulness to Meta's social graph.
  • A new monetization path: Conversations shape ads and feeds, Meta One sells extra AI capacity, Meta Business Agent charges businesses per token or by subscription, and the Muse agent adds separate subscriptions, giving Meta several ways to recover its infrastructure spending.
  • Platform power under scrutiny: The WhatsApp chatbot ban and the European Commission's interim order show regulators treating messaging apps as gateways for AI assistants, not just communication tools.
  • Safety at population scale: Teen interactions with AI characters, an account-support chatbot that could be talked into handing over accounts, and an image feature built on opt-out consent show how small design choices become large harms when an assistant serves billions of people.

Meta AI Limitations and Open Questions

Meta AI's scale magnifies its weak points as much as its strengths. Several of the issues below are matters of policy and product design rather than model quality, so they can change quickly with a regulatory decision or a new default setting. The main limitations and open questions as of late October 2026 are:

  • Data use without opt-out: Outside the regions initially excluded, people cannot use Meta AI without their conversations becoming a signal for feed and ad personalization, although sensitive topics are excluded; only Incognito Chat on WhatsApp and the Meta AI app offers a conversation Meta says it cannot read.
  • Uneven availability: New Meta AI features, including the Muse Spark upgrade and many shopping and voice features, launched first in the United States and Canada, and some regions with strict privacy rules received them later or not at all.
  • Reliability: Muse Spark is a strong model by independent measures, but it can still be wrong, and health and shopping answers should be checked against authoritative sources; independent testers also found gaps between benchmark claims and hands-on results.
  • Consent design: The withdrawn Muse Image feature showed the limits of opt-out defaults for features that use other people's photos, and images generated before withdrawal were not automatically deleted.
  • Agents with real authority: The Instagram support-chatbot incident showed that giving an AI agent the power to change account ownership without strong identity checks invites abuse, a lesson that applies to every agent Meta deploys.
  • Unresolved regulation: The European Commission's WhatsApp investigation, inquiries in Italy and Brazil, and US scrutiny of companion chatbots for minors were all unresolved as of October 2026, leaving basic questions of AI governance for consumer assistants open.

Meta AI History

  • September 2023: Meta introduced Meta AI, built on Llama 2, along with celebrity-voiced AI personas, across Messenger, Instagram, and WhatsApp, and on Ray-Ban Meta glasses.
  • 2024: Meta AI moved to newer Llama models and gained the meta.ai website and wider placement in app search bars, and in July Meta opened AI Studio so users and creators could build their own AI characters.
  • April 29, 2025: Meta launched the standalone Meta AI app, built with Llama 4 and replacing the Meta View glasses app, with memory, voice conversations, and a public Discover feed.
  • August 2025: Reuters reported on Meta's internal "GenAI: Content Risk Standards," prompting a US Senate inquiry and changes to Meta's chatbot rules.
  • September 25, 2025: Meta added Vibes, a feed of AI-generated short videos, to the Meta AI app and meta.ai.
  • October 2025: Meta changed its WhatsApp Business policy to bar general-purpose AI assistants, announced that Meta AI conversations would inform ads and feeds, and set out new teen AI protections.
  • December 2025: The European Commission opened an antitrust investigation into the WhatsApp policy on December 4, the ad-personalization change took effect on December 16, and Meta announced its acquisition of Manus on December 29.
  • January 2026: The WhatsApp ban on rival general-purpose assistants took full effect on January 15, and on January 23 Meta paused teens' access to AI characters worldwide.
  • February to March 2026: The Commission sent Meta a Statement of Objections, Meta confirmed it was testing Vibes as a standalone app, and in March Meta offered rival assistants paid access to WhatsApp in Europe and launched its Meta AI Support Assistant, while Meta-owned Manus launched its My Computer desktop agent on March 16.
  • April 8, 2026: Meta Superintelligence Labs released Muse Spark, and Meta relaunched the Meta AI app and meta.ai with Instant and Thinking modes.
  • April 27, 2026: China's National Development and Reform Commission prohibited Meta's acquisition of Manus and ordered the parties to withdraw from the deal.
  • May 2026: On May 12, Meta brought Muse Spark's voice, Live AI, and Marketplace shopping features to more surfaces and began rolling it out to Ray-Ban Meta and Oakley Meta glasses; on May 13, WhatsApp introduced Incognito Chat.
  • June 2026: Hijacked Instagram accounts revealed the support-assistant flaw, which Meta patched around June 1; Meta launched Meta Business Agent on June 3; and the Commission ordered free WhatsApp access for rival assistants on June 9, while Meta moved to cut ties with Manus.
  • July 2026: Meta launched Muse Image on July 7 and withdrew its Instagram @-mention feature days later, released Muse Spark 1.1 to API customers on July 16, and reported a 60 percent increase in daily Meta AI users since the Muse Spark rebuild.
  • August to September 2026: Muse Spark 1.2 (August 5) and 1.3 (September 2) followed; Manus, which announced its separation from Meta on August 11, resumed independent operations on September 1; Meta launched the Muse agent on September 8 and Meta One worldwide on September 15; and at Connect on September 23 it introduced new AI glasses and announced Muse access from them.
  • October 2026: Meta released the Muse Wonder models on October 12 and the open-weight Muse Spark Open later that month, extending the Muse family beyond Muse Spark.

Manus

Manus is an autonomous, general-purpose AI agent designed to plan, execute, inspect, and deliver the results of complex computer-based tasks rather than only returning text like a conventional chatbot. Its documented operating environment includes a virtual computer with internet access, persistent file storage, software installation, a command-line interface, browser access, and connections to external services, and since March 2026 its My Computer feature has also let it work directly on a user's own Mac or Windows PC. Manus was built by Butterfly Effect, a company founded in Beijing in 2022 whose team had earlier released the Monica browser assistant, and it entered invitation-only beta on March 6, 2025, before the company moved its headquarters to Singapore in mid-2025. Meta acquired Manus at the end of 2025, but Chinese regulators ordered the deal unwound in April 2026, and Manus resumed operating as an independent company on September 1, 2026.

Manus represents one approach to agentic AI, in which a user delegates an outcome rather than specifying every intermediate action. Depending on the task, Manus can perform multi-source research, analyze datasets, write and execute programs, build and publish full-stack websites, prepare documents and presentations, edit media, create games, and automate recurring workflows. The system is intended to combine foundation models with tools, persistent workspaces, and computer-use capabilities so that users can obtain finished artifacts or working applications instead of only suggestions or code fragments.

The distinction between Manus and a conventional chatbot is operational rather than absolute. Manus still relies on language models for interpretation, planning, and generation, but it places those models inside a runtime that can open files, use a browser, run code, call services, observe the results, and revise its work. This design can reduce the amount of manual coordination required from users, although it does not eliminate the need for review, permissions, testing, or human judgment in high-impact situations.


Manus is built around the idea that an AI system should not only explain what to do, but also use a computer, coordinate tools, and complete the work. Its central value is the conversion of a natural-language objective into an end-to-end result.


Manus and Meta

Meta announced on December 29, 2025, that it would acquire Manus in a deal reported at more than $2 billion, when the service had an annual revenue run rate of about $125 million and said it had processed more than 147 trillion tokens and created more than 80 million virtual computers. Meta said Manus would keep selling its subscription service while its team helped bring agent technology into Meta AI and Meta's business products, and it promised that there would be no continuing Chinese ownership interests in Manus and that Manus would end its services in China. By March 2026, roughly 100 Manus employees were working from Meta's Singapore offices and chief executive Xiao Hong was reporting to Meta chief operating officer Javier Olivan. During that period Manus kept shipping under Meta ownership, including the My Computer desktop agent described below.

Beijing objected to the deal from the start. China's Ministry of Commerce said in January 2026 that it would assess whether the acquisition complied with Chinese laws and regulations, and co-founders Xiao Hong and Ji Yichao were reportedly barred from leaving China from March while the review continued. On April 27, 2026, the National Development and Reform Commission, China's top planning agency, issued a one-line statement prohibiting the foreign acquisition of Manus and requiring the parties to withdraw from the transaction, reflecting concern that technology developed in China could be transferred abroad. Meta said the deal had complied fully with applicable law, but in June 2026 it moved to cut ties, separating Manus's operations and data from its own systems while Manus's founding investors, led by Tencent, HSG (formerly Sequoia China), and ZhenFund, arranged to buy the company back at roughly the price Meta had paid.

On August 11, 2026, Manus told users that it would soon return to operating as an independent company. To comply with regulatory requirements in specific jurisdictions, it deleted data that certain users had generated on or after December 29, 2025, during an August 23 to 24 window (Singapore time), after giving affected users time to back up their data and offering restoration from August 25; the company said the step was not the result of a security incident. Manus formally resumed independent operations on September 1, 2026, with founders Xiao Hong, Ji Yichao, and Zhang Tao still in charge, and on October 8 its parent Butterfly Effect said it had raised more than $500 million in a round co-led by Boyu Capital and IDG Capital, with Tencent, HSG, and ZhenFund also investing. The episode left Meta to pursue personal agents through its own Muse agent, which launched a week after Manus regained its independence, and it has become a reference case for how far Chinese authorities will go to control AI technology built by Chinese founders who move offshore.

Manus Mechanisms

When a user submits a request in natural language, Manus interprets the desired outcome, constraints, and available resources before constructing a multi-step execution plan. The plan may include research, file operations, code generation, browser actions, data transformation, media production, testing, and packaging. The system can revise the plan when observations from a tool call change the requirements or reveal an error. This makes the agent a stateful workflow executor rather than a single-pass text generator.

An orchestration harness coordinates the model calls and tools needed to carry out the plan. Depending on the assignment, Manus can use a shell, Python environment, graphical or headless browser, document parser, spreadsheet tools, image and video generation systems, databases, and third-party application programming interfaces. The agent can also use browser automation to navigate dynamic websites, although authentication challenges, anti-bot systems, and sensitive actions may require a user to intervene.

Manus maintains task state through project files, intermediate artifacts, execution traces, and context-management techniques. Its official documentation describes a sandboxed virtual computer with internet access, persistent storage, and the ability to install software or create custom tools. In practice, the resulting deliverable may be a downloadable file, a source-code repository, a deployed website, a data analysis, a media project, or a long-running workflow.

Manus Principles

The system rests on four related principles: high-level task delegation, composable tool use, iterative verification, and bounded human control. Delegation allows a user to specify the objective, scope, constraints, and acceptance criteria while Manus handles many intermediate steps. This changes the human role from manually operating every application to defining the work and reviewing the results.

Tool use connects language-based reasoning with deterministic software execution. A language model can propose a query, script, design, or action, while a browser, compiler, database, parser, or media editor produces an observable result. This division is important because the model's output is not necessarily the final authority; the result of executing and inspecting the proposed action provides additional evidence.

Verification is implemented through repeated observation and correction. Manus may inspect compiler errors, compare a rendered page with requirements, validate a data schema, run tests, check a generated file, or revise a design after viewing the result. These practices are related to AI verification and validation and model evaluation, but successful execution does not guarantee that the work is factually correct, secure, legally appropriate, or aligned with the user's unstated preferences.

Human control remains important because an agent can have access to private data, authenticated accounts, local files, production systems, or external communication channels. Users determine which connectors, folders, websites, and actions are authorized, and should require confirmation before irreversible operations such as financial transactions, public publishing, bulk deletion, or changes to production infrastructure. Manus builds this principle into its products through approval prompts, folder-scoped access on local computers, and the option for a user to take over the browser when a step needs a human decision.

Manus Architecture and Technical Mechanisms

Manus is a compound AI system rather than a standalone language model. It combines one or more large language models with an orchestration layer, software tools, persistent storage, browser and computer-use interfaces, and verification routines. Frameworks such as LangChain provide libraries that developers can use to assemble similar components in their own applications, whereas Manus provides a managed product environment in which planning, tool invocation, file management, and artifact delivery are integrated.

The system can route different steps to different model capabilities. A higher-reasoning model may be used for decomposition, code synthesis, or difficult decisions, while faster models may handle routing, summarization, formatting, or routine transformations. Context management is necessary because long-running work can produce large volumes of terminal output, webpage content, screenshots, source code, and file diffs. Techniques associated with attention mechanisms, summarization, retrieval, and state tracking help determine which information remains active, although the resulting context can still be incomplete or misleading.

The execution lifecycle resembles a perception, planning, action, and observation loop. A planner converts the request into a task graph or checklist, a coordinator dispatches individual steps, and execution tools perform those steps. An observation stage then captures standard output, error messages, screenshots, page states, test results, or changed files. The agent uses these observations to decide whether to continue, revise its approach, request user input, or stop.

Error recovery is a central part of the design. If a program raises an exception, a website changes its layout, a package fails to install, or a generated artifact does not satisfy a requirement, Manus can diagnose the available evidence and attempt a correction. Recovery remains probabilistic, and repeated retries can waste time or obscure the original problem. A user should therefore inspect important outputs and define stopping conditions for workflows that can affect external systems.

Cloud Computer environments provide isolated project workspaces with a Linux operating system, files, network access, package management, and the ability to run local servers or development tools. Persistent workspaces allow a project to be paused, inspected, edited manually, and resumed later. Isolation reduces the blast radius of arbitrary code execution, but it is not a substitute for least-privilege permissions, dependency review, network controls, secrets management, or testing.

AI Offerings, Agent Harnesses, and Applications

Manus's product ecosystem includes the core agent, cloud execution environments, creation tools, recurring automations, integrations, local computer use, and the Cue personal-agent application. The offerings are connected by the idea that an agent should be able to move from an instruction to an observable result. Some capabilities are general-purpose, while others are specialized environments intended to keep the agent's context and tools focused on a particular type of work.

Manus Agent Generations and Cascade

The principal agent has evolved through releases including Manus 1.0 in March 2025, Manus 1.5 in October 2025, and Manus 2.0 on September 28, 2026. Manus 1.5 was a re-architected system with full and cost-efficient variants that emphasized speed, reliability, larger tasks, collaboration, a central artifact library, and full-stack application development with persistent backends, databases, user authentication, embedded AI capabilities, custom domains, analytics, and version control. Manus 2.0 introduced Cascade, which Manus describes as an in-house agent harness that keeps a project lean and brings in specialized capabilities only when a task requires them. These were company-reported capabilities, and the performance figures Manus published with each release were internal results rather than independent benchmark scores.

Cascade is intended to load specialized capabilities when the task requires them instead of carrying every tool and context element into every run. In a research project, this may mean coordinating browsing, note-taking, file creation, and synthesis. In a software project, it may mean coordinating code editing, database changes, testing, deployment, and documentation. Manus reported that, in one tested configuration, Cascade used 23.2 percent fewer tokens, completed tasks 28.2 percent faster, and cost 32 percent less than the previous system, figures it presented as results for a particular configuration rather than a general evaluation.

Cloud Computer, Browser, Wide Research, and Mail

Cloud Computer is a persistent virtual environment for projects that need a server, files, installed packages, or scheduled execution to remain available beyond a single session. A project can use it to host a development server, maintain a database, run a bot, process files, or provide a persistent home for an Automation, and Manus 2.0 made dedicated Cloud Computers available for purchase. The environment is separate from the user's ordinary local computer, subject to the product's permissions, pricing, and operational limits, and Manus's pricing page in October 2026 listed individual plans at $20, $40, and $200 a month with 4,000, 8,000, and 40,000 credits.

Cloud Browser and related browser-control features allow Manus to navigate websites, fill forms, extract information, and interact with dynamic web applications. Browser actions can fail when a site requires a CAPTCHA, multifactor authentication, an unavailable API, or a human decision involving sensitive information. In those cases, a user may need to take control of the browser or complete the step manually.

Wide Research is a parallel research mode that divides a broad assignment among multiple research processes before synthesizing their findings. Parallel investigation can improve coverage and reduce the time required for independent searches, but it does not automatically guarantee source quality, non-duplication, or correct synthesis. Research outputs still require source verification, especially when the topic is current, controversial, or high stakes.

Mail Manus provides an email-based way to submit instructions, attachments, or correspondence for agent processing. Email delegation can be useful for asynchronous work, but it also expands the security boundary. Organizations should consider sender authentication, attachment handling, confidential information, retention, and the possibility that an incoming message contains a prompt injection attack.

Manus Studio and Creation Environments

Manus Studio is a shared workspace where people and agents create, inspect, revise, and publish artifacts. With Manus 2.0 it replaced the earlier Manus Desktop app, and Manus describes it as a desktop application and project workspace that brings in only the tools a project calls for, keeping the agent's context focused on the work at hand. The release describes eight broad environments:

  • Full-stack web and mobile development: Manus can scaffold applications, configure backend services, add persistent data storage and authentication, connect services, test through a browser, and publish supported projects. Development still requires code review, security testing, accessibility checks, dependency management, and operational monitoring.
  • Documents and PDFs: The document environment can produce structured reports, procedures, business briefs, and other formatted files. Users should verify quotations, calculations, citations, page layout, and any legal or regulatory claims.
  • Spreadsheets and data visualization: Manus can clean datasets, write formulas, perform analysis, and create charts or dashboards. Results depend on data quality, correct assumptions, appropriate statistical methods, and inspection of the generated formulas and transformations.
  • Slides and presentations: The system can turn research, business metrics, or an outline into a visual presentation with charts and speaker notes. Human review remains necessary for narrative accuracy, visual hierarchy, accessibility, and the interpretation of quantitative evidence.
  • Video Editor: The timeline-based editor exposes clips, images, text, motion graphics, and audio as separate components so that a user can revise individual elements rather than regenerate an entire video, and an Alchemy mode starts from an idea and lets Manus handle creative direction. Manus describes use cases including short product advertisements, AI-generated user-generated content, tutorials, vlogs, and motion graphics.
  • Game Dev: The game environment combines code, assets, animation, audio, level design, and publishing. Manus describes support for playable templates, a manual edit panel, web publishing, and multiplayer projects backed by a persistent Cloud Computer, although production games still require testing for performance, security, networking, copyright, and moderation, as discussed in Game Development with Generative AI.
  • Research and knowledge work: Broad research workflows can combine browsing, source collection, data analysis, writing, and artifact production. The quality of the result depends on source selection and the agent's ability to distinguish evidence from speculation.
  • Automation and operations: Projects can combine connectors, schedules, event triggers, files, and business logic to support recurring work. Such workflows should have clear ownership, logging, failure notifications, and a controlled rollback path.

Automations, Scheduled Tasks, and Connectors

Manus provides scheduled work and event-triggered Automations for recurring operations. A scheduled task can run at a defined time or interval, while an event-triggered workflow can begin after a change in a connected service, such as a new email, a calendar event, a Slack message, an updated Notion page, or a change in advertising data. This model extends scheduled workflows from time-based execution toward event-driven operations, and Manus says an Automation can be set up with a single prompt.

Connectors provide permission-scoped access to external platforms. Manus identifies integrations with services such as Shopify, Slack, Google Workspace, Meta Ads, Instagram, Notion, Airtable, Similarweb, Zoom, HubSpot, Canva, and Supabase. The practical security of a connector depends on the permissions granted, how credentials are stored, how data is logged, and whether actions require approval.

Brand Skills and related reusable instruction packages can encode an organization's tone, visual identity, product information, and operating procedures. They can improve consistency across storefront updates, marketing materials, reports, and customer communications, but they do not replace review of claims, brand compliance, copyright, or regulatory requirements. The Manus API provides a programmatic route for triggering agent work, managing projects, and retrieving artifacts from external software.

Cue Personal Agent

Cue is a separate Manus application for personal agents on phones and desktop computers, introduced alongside Manus 2.0 and built on the same foundation. Manus describes Cue as a way to give an agent a task, allow it to work on its own computer, and coordinate several agents around a shared goal in group chats where they hand work to one another. Each Cue agent has its own email address, phone number, wallet with a user-defined budget, and computer, so it can send messages, take calls, pay for things within its budget, and handle everyday services such as ordering at a restaurant by QR code.

These capabilities expand the usefulness of agents beyond document and software production, but they also create a higher-risk permission model. Communication, payments, calls, and bookings require explicit limits, transaction review, identity protection, and clear records of what the agent did. Cue launched in early access as a free service requiring an invite code, with an iOS app awaiting App Store review at launch, and its availability, features, and regional support may change.

Manus My Computer

My Computer brings Manus out of its cloud sandbox and onto a user's own computer. Manus introduced it on March 16, 2026, while the company was owned by Meta, as the core capability of the new Manus Desktop app, which Manus 2.0 later upgraded into Manus Studio. Instead of uploading every local file to the cloud or manually copying commands, a user connects chosen folders and lets Manus run command-line instructions in the computer's terminal to read, analyze, edit, and create files and to launch and control local applications. Manus framed the launch as moving the agent from its own workspace into the user's, and it arrived during a wave of locally running agents that included OpenClaw, Perplexity Computer, and Anthropic's Claude Cowork.

How My Computer Works

My Computer is built on folder-scoped access. A user adds folders under My Computer in the desktop app, authorizes them, and can turn on Work locally so that a chosen folder becomes a task's workspace; Manus can ask for access to another folder mid-task, and disconnecting a folder does not delete its files. Planning still happens in the cloud, so file content Manus reads becomes part of the task, and local work draws on the same Manus credits as cloud tasks. Its capabilities fall into five broad groups:

  • Local file and directory management: The agent can inspect approved folders, reorganize files, batch-rename items, convert formats, identify duplicates, or process documents and images. In Manus's launch examples, a florist had thousands of mixed shop photos sorted into categorized folders, and an accountant had hundreds of invoices renamed to a standard format.
  • Terminal and command-line execution: The agent can run build scripts, manage Git repositories, install packages, execute tests, and use local development tools such as Python, Node.js, Swift, and Xcode. Manus said one colleague used it to build a real-time meeting translation and subtitle app for the Mac in about 20 minutes without opening Xcode.
  • Local hardware: Manus can use the computer's processor and GPU to train machine learning models or run large language model inference, turning an idle desktop or Mac mini into an always-available assistant.
  • Cross-application workflow automation: The agent can coordinate actions across approved desktop programs, combine local files with connectors such as Gmail and Google Calendar, and produce a final artifact in a designated location. Paired with Projects and Automations, it can run recurring local routines, such as tidying a Downloads folder every morning or building a weekly report from local data.
  • Computer Use and Remote Control: Since Manus 2.0, Manus can operate approved apps and the browser in a visible workspace on the user's computer. Remote Control lets a user send voice or text instructions from the Manus mobile app and watch the desktop live, so a person away from home can, for example, have Manus find a contract on a home computer and email it to a client.

My Computer Permissions and Safety

At launch, every terminal command required explicit approval, with an Allow Once option for reviewing each operation and an Always Allow option for trusted tasks. Manus's current documentation describes three permission modes: Ask for approval, which asks before every local action; Approve for me, the default, which auto-approves actions it judges safe and blocks high-risk ones such as sending private keys to the internet; and Full access, which runs actions without approval after a warning. Even in the approval modes, network access, changes outside the workspace, switching the working folder, and granting full access to a single command still require the user's consent, and on Windows the approval modes need a one-time setup step. Computer Use asks separately the first time it needs an app, with Always allow and Never choices that apply to all agents on that computer.

The documentation is candid about the limits of this model. Approval modes restrict writes rather than reads, so commands can read any file the user's account can open, Full access runs commands with the user's own permissions, and administrator rights still require approval. Manus recommends choosing the smallest workspace folder a task needs rather than a whole home folder, using Ask for approval for sensitive files, reading each request before approving it, and backing up with Git or a copy before bulk edits. More broadly, users should separate personal data from project data, protect credentials and secrets, and require confirmation before network pushes or external communications, because local execution makes a mistaken or manipulated action more consequential than the same action in a disposable cloud sandbox.

My Computer Availability

My Computer and Computer Use require the Manus Studio desktop app on Apple Silicon Macs or on Windows 10 and 11; Intel Macs and Linux are not supported, and the web version of Manus does not include them. Remote work requires the computer to stay plugged in, awake, signed in, and running Studio, tasks stop if it sleeps or shuts down, and a task may temporarily switch to the cloud if a feature fails locally, so Manus points users who need always-on work, such as bots or servers, to a Cloud Computer instead. Because My Computer launched under Meta ownership and continued after the separation, it is the Manus capability that most directly parallels Meta's own Muse agent, which runs in Meta-managed cloud virtual machines and reaches local Apple apps only through its macOS desktop app.

Why Manus Matters

Traditional digital work often requires people to learn many disconnected interfaces, translate information between applications, and perform repetitive administrative actions. Conventional robotic process automation can handle rigid, predictable sequences, but it is less effective when a workflow contains unstructured documents, changing websites, ambiguous instructions, or content creation. An autonomous general agent attempts to combine flexible reasoning with cross-application execution.

This approach can allow a non-programmer to produce a custom internal tool, analyze a business dataset, prepare a research report, or automate a recurring process without assembling a separate team for every stage. It can also help technical users by handling routine investigation, environment setup, testing, documentation, and coordination across files and services. The practical value depends on whether the agent completes the task reliably enough to offset the cost of supervision and correction.

The same capabilities create risks. Long workflows can accumulate small errors, models can misunderstand requirements, and a browser agent can encounter malicious instructions in webpages, emails, or documents. Other concerns include privacy, access control, data retention, cloud costs, software supply-chain risk, inaccurate research, and the possibility that a generated application will be deployed without adequate testing. Manus therefore belongs alongside, rather than above, established practices for privacy, cybersecurity, software engineering, data governance, and human review.

Manus for Enterprise and Small Business

Small and medium-sized businesses can use Manus as a general operations layer for research, content, analytics, software creation, and workflow automation. An e-commerce business might combine Shopify data, advertising information, competitor research, product catalogs, and spreadsheets to produce a pricing or inventory report. A service business might build a custom booking portal, invoice tracker, customer-support workflow, or internal dashboard rather than buying a separate application for each narrow task.

Full-stack development can be useful when an organization needs a small internal application that does not fit an existing software product. A natural-language interface can help a non-specialist describe a database schema, dashboard, workflow, or permission rule, while the agent creates and tests the implementation. The resulting software still needs conventional engineering controls, including authentication review, authorization testing, backups, monitoring, data retention rules, and maintenance ownership.

For marketing and consumer packaged goods, an agent can help prepare catalogs, analyze sales data, draft campaign materials, monitor competitors, and produce recurring executive summaries. These uses combine data science, databases, spreadsheets, web research, and content generation. The main risk is that a polished report or dashboard can conceal incorrect inputs, weak assumptions, or unsupported conclusions, so decision-makers should retain access to underlying data and calculations.

Manus Technical Limitations

Autonomous agents are vulnerable to compounding errors. A mistaken interpretation of a requirement can produce the wrong research question, which can lead to an incorrect dataset, an inappropriate implementation, and a polished but unusable final artifact. Long-horizon workflows also consume more time and model calls than ordinary chat requests, and failures can occur at any boundary between the model, a browser, an API, a file, or an external service.

Browser and computer-use systems face additional limitations. Websites can change their layouts, block automation, require human verification, or expose content that is difficult to interpret reliably. A system may also appear to have completed a step when it has only changed a page visually without achieving the intended backend result. Testing through independent checks, APIs, logs, or human inspection is therefore important for consequential tasks.

Benchmark scores should be interpreted cautiously. GAIA, a benchmark developed with researchers at Meta, evaluates real-world assistant tasks that involve browsing, multimodal inputs, reasoning, and tool use, but performance on a benchmark does not establish safe or reliable operation across all production workflows. Product demonstrations and internal performance claims are useful evidence about intended capabilities, while independent replication, task-specific evaluation, and operational monitoring provide stronger evidence about real-world reliability.

Manus Security, Privacy, and Governance

An agent that can read private files, execute code, use a browser, and access authenticated services has a broader attack surface than a text-only chatbot. The principal risks include excessive permissions, credential exposure, malicious downloads, insecure generated code, data leakage, unauthorized external actions, and inadequate separation between untrusted content and trusted instructions. Organizations should apply least privilege, isolate environments, protect secrets, maintain logs, and define approval requirements for high-impact actions.

Indirect prompt injection is a major risk for browser and document agents. An attacker can place instructions in a webpage, email, attachment, advertisement, image, or document that the agent later processes. If the agent treats that content as an instruction rather than as untrusted data, it may disclose information, follow an attacker-controlled workflow, or invoke a connected tool. Research on computer-use agents has identified risks including visual clickjacking, indirect prompt injection leading to chained tool execution, and manipulation of the agent's intermediate reasoning or interface context.

Defenses include separating instructions from observations, tracking the provenance of inputs, limiting which tools can be called from untrusted content, using allowlists and scoped credentials, requiring confirmation for sensitive operations, and testing agents with adversarial content. Security researchers emphasize that prompt-injection defenses are improving but the problem is not solved. A browser agent should therefore be treated as operating in an adversarial environment, not as an infallible personal operator.

Governance also includes accuracy, accountability, and data lifecycle management. Users should know which model or service processed their information, where outputs are stored, how long artifacts and logs are retained, and who is responsible for correcting an erroneous action. Manus's separation from Meta showed how quickly these questions can become practical, when regulatory requirements led to the deletion of some users' data created during the Meta period. For high-impact domains, agent outputs should remain subject to qualified human review, documented procedures, and applicable legal and organizational controls.

Manus History

  • 2022 to 2023: Xiao Hong founded Butterfly Effect in Beijing in 2022, and in 2023 the company released Monica, a browser assistant that combined several commercial language models for translation, summarization, and writing.
  • March 6, 2025: Manus entered invitation-only beta and drew wide attention as an example of an agent that could carry out asynchronous, multi-step work in a cloud environment.
  • Mid-2025: After a funding round led by Benchmark, Butterfly Effect moved its headquarters to Singapore, relocating part of its China-based staff and laying off the rest.
  • October 2025: Manus released Manus 1.5, a re-architected agent with full and cost-efficient variants and full-stack web application development.
  • December 29, 2025: Meta announced its acquisition of Manus in a deal reported at more than $2 billion.
  • January to March 2026: China's Ministry of Commerce opened a review of the deal, co-founders Xiao Hong and Ji Yichao were reportedly barred from leaving China, and on March 16 Manus launched My Computer in the new Manus Desktop app for macOS and Windows.
  • April 27, 2026: China's National Development and Reform Commission prohibited the foreign acquisition of Manus and ordered the parties to withdraw from the deal.
  • June 2026: Meta moved to cut ties with Manus, separating operations and data, while Manus's founding investors arranged to buy the company back.
  • August 2026: On August 11 Manus announced its return to independent operations, and on August 23 to 24 it deleted data that certain users had created since December 29, 2025.
  • September 2026: Manus resumed independent operations on September 1 and released Manus 2.0 on September 28 with Cascade, Manus Studio, Cloud Computer, Automations, Video Editor, Game Dev, Remote Control, and the Cue personal-agent app.
  • October 8, 2026: Butterfly Effect announced that it had raised more than $500 million in a round co-led by Boyu Capital and IDG Capital, its first since the Meta deal was unwound.

Meta Muse


Muse is Meta's first consumer AI agent, launched in the United States on September 8, 2026, after development under the internal codename "Hatch." Where Meta AI answers questions in a standard conversational turn, Muse executes multi-step tasks autonomously: you tell it what needs to get done, and it performs web automation, shops online, fills out forms, plans trips, sends emails, schedules events, and can even negotiate a recurring bill on your behalf. Meta pitches it as an accessible agent for everyone, with no technical experience needed and no steep learning curve. Behind its simple chat interface, the system dynamically writes code or sets up its own connections to external services with public APIs in order to finish complex, long-horizon workflows. Muse keeps working asynchronously in the cloud after the app is closed, returns notifications when something changes or when an action needs explicit human approval, and suggests proactive ideas of its own based on what it has learned about the user over time.

Muse operates as a separate product from the standard Meta AI app, with dedicated iOS and Android applications, a web interface at muse.ai, a native macOS desktop app added on September 17, 2026, and a chat entry point inside WhatsApp. Support for Meta's AI smart glasses via a dedicated wake word and a standalone pocket hardware device called Muse Charm were also announced at Meta Connect for release later in 2026. To use the service, a person needs a Meta account (though not necessarily a Facebook or Instagram profile) and must be 18 or older. After launching in the United States, Muse expanded rapidly to Canada on September 18, reached Japan on October 23 as its first market outside North America, and launched in Singapore with localized service connectors on October 27, 2026.

The agent runs on Meta's proprietary Muse family of models developed by Meta Superintelligence Labs (MSL), beginning with Muse Spark 1.3 at launch and expanding to include the frontier Muse Wonder models for paid subscribers and the terminal-based Muse Code agent for software engineers. Consumer access to Muse is free for most everyday tasks up to a weekly token limit, while two paid subscription tiers, Power at $20 a month and Maximum at $100 a month, add higher compute capacity and priority access to new model weights. As Meta's first paid consumer AI product, Muse carries no advertising and keeps agent conversations out of Meta's ad-targeting systems, with the company expecting to generate long-term revenue through subscriptions and small merchant fees on agent-driven transactions.

Muse differs from earlier digital assistants in two fundamental ways that define the current wave of autonomous software. Unlike a conversational chatbot, it holds persistent, long-term goals and acts across connected third-party services rather than producing a single text answer; unlike self-hosted open-source agents such as OpenClaw, which directly inspired its architecture, it runs entirely on cloud virtual machines that Meta provisions and secures, so ordinary users never have to install local runtimes or manage API keys themselves. That central trade-off, gaining effortless convenience in exchange for trusting Meta with personal email, calendars, payment methods, and browsing sessions, forms the core tension in nearly every technical and privacy assessment of the platform.


Muse is Meta's bet on "personal superintelligence": AI directed by each individual toward their own goals, rather than centrally automating all work.


Mechanisms

Each Muse agent works inside its own Muse Secure VM, a dedicated cloud virtual machine on Meta's infrastructure that houses the agent runtime, a full headless and viewable web browser, the user's workspace data, and any stored integration state. Users can watch the agent work live in its built-in browser window, and strict virtualization boundaries ensure that no other user's agent can reach the machine. Inside the virtual machine, the agent runs in a sealed runtime cell that is engineered under the assumption that it will regularly ingest untrusted web data, while privileged services outside that cell hold actual passwords and enforce security policy. Credentials are injected just in time at the network boundary by swapping a surrogate token for the real secret at the exact moment an authorized request leaves the environment, meaning the language model itself never holds a plaintext password or API key that could be exfiltrated.

A second, independent system component called Sentinel runs on the same virtual machine image but is strictly isolated from the Muse reasoning loop at the operating-system level. Sentinel serves as the sole permission authority for external connector actions and for any network request that attempts to leave the virtual machine: Muse proposes a tool call or outbound connection, and Sentinel evaluates the request against the user's configured policy to answer allow, deny, or ask. When Sentinel requires human confirmation, agent execution halts immediately and the approval prompt appears as an operating-system-level dialog inside the Muse application rather than as a conversational turn, ensuring that malicious text hidden on a web page cannot spoof a user's consent. Meta reinforces this dual-agent architecture with five overlapping layers of defense against prompt injection:

  • Model training: the underlying Muse models undergo adversarial alignment to recognize and resist untrusted instructions embedded in retrieved content.
  • Harness marking: every piece of external data entering the context window from web pages, emails, or documents is tagged with untrusted-source provenance markers before the model processes it.
  • Classifiers: independent safety models continuously scan page text, images, and file downloads for hidden injection payloads and block outbound traffic if the agent attempts to transmit personal data to an unapproved destination.
  • Deterministic checks: traditional procedural code validates tool arguments and enforces read/write permission boundaries regardless of what the neural model decides to output.
  • Credential separation: because the reasoning environment only handles surrogate tokens and never sees raw passwords or full payment card numbers, a compromised context window holds no reusable secrets to steal.

Payments and e-commerce checkouts follow the same principle of zero-trust credential isolation. At launch, Muse executed purchases through Link by Stripe, whose agentic wallet issues a single-use virtual card locked to the specific merchant and approved amount, making Muse the first AI agent covered by Link's consumer purchase protections. On September 21, 2026, Shopify expanded this ecosystem by launching native agentic checkout through Shop Pay across all Shopify merchants, PayPal added merchant checkout support the following day, and Meta announced upcoming 1Password integration to let Muse authenticate with logins a person already stores in their personal vault. Across every payment rail, Sentinel pauses execution to require explicit user confirmation before any money is spent, and the app maintains a permanent audit trail of past and scheduled actions.

Muse's day-to-day usefulness comes from its persistent memory architecture and proactive goal management. The system remembers personal details a user mentions in passing, consolidates what it has learned through periodic background memory compaction, breaks multi-week objectives into structured action plans, and surfaces timely task suggestions in a dedicated Ideas tab while allowing users to inspect, edit, or delete stored facts at any time. The core intelligence driving these workflows is Muse Spark, a natively multimodal reasoning model introduced by Meta Superintelligence Labs on April 8, 2026, and upgraded to Muse Spark 1.3 just before launch. Muse Spark supports native tool use, visual chain-of-thought reasoning over browser screenshots, and multi-agent orchestration, including a high-compute "Contemplating" mode in which several specialized sub-agents reason in parallel before synthesizing a final plan, while the newer Muse Wonder model family brings expanded reasoning depth to paid subscribers.

Muse Code

Alongside the consumer Muse assistant, Meta Superintelligence Labs developed Muse Code, a specialized terminal-based coding agent built for software engineers and technical teams. First introduced in August 2026 alongside the Muse Spark 1.2 model release, Muse Code operates directly inside a developer's local command-line interface and repository environment, competing directly with command-line engineering tools such as Anthropic's Claude Code, OpenAI's Codex CLI, and open-source terminal harnesses. Rather than acting as a simple autocomplete plugin inside an editor, Muse Code functions as an autonomous AI pair programmer capable of navigating multi-file codebases, reading documentation, executing shell commands, running test suites, debugging compiler or runtime errors, and preparing git commits across complex software projects.

In October 2026, Muse Code became the primary deployment vehicle for Meta's highest-compute frontier model, Muse Wonder Full. When Meta unveiled the Muse Wonder model family on October 12, 2026, it split the release into two distinct tiers: the standard Muse Wonder model was rolled out to consumer Muse Power and Maximum subscribers to handle everyday agentic reasoning, while the heavier, unpruned Muse Wonder Full model was reserved exclusively for Muse Code subscribers. This separation reflects the distinct computational demands of repository-scale software engineering, where an agent must maintain coherent reasoning across hundreds of thousands of tokens of source code, dependency graphs, and terminal execution logs without losing track of architectural constraints.

Within Meta's broader developer ecosystem, Muse Code bridges proprietary frontier capabilities and open-weight local deployment. Developers who want hosted, frontier-grade autonomous coding can subscribe to Muse Code to access Muse Wonder Full and Muse Spark 1.3 with managed tool-use harnesses. Teams that require strict on-premises code privacy or custom fine-tuning can instead pair local terminal agent frameworks with Meta's open-weight releases, including the lightweight 30-billion-parameter Muse Glimmer model (released under the Apache 2.0 license in August 2026) and Muse Spark Open, an 816-billion-parameter (42-billion active) mixture-of-experts model released in October 2026 with a 1-million-token context window capable of holding entire codebases in active memory.

The Physics

A personal agent that holds a person's real accounts and executes live transactions must operate under hard constraints of security, platform trust, and compute capacity. Meta's technical documentation is unusually candid about several of these boundaries, and independent security researchers and web infrastructure providers have documented others since launch. The most important operational constraints include:

  • Residual injection risk: Meta's safety documentation states that Muse Spark 1.3 remains susceptible to adaptive jailbreaks and indirect prompt injection in complex agentic settings, which is why the Secure VM architecture assumes the language model will occasionally be compromised. Meta backs its containment design with a public bug bounty paying up to $300,000, including up to $130,000 for a reproducible prompt injection that exfiltrates data or triggers unauthorized actions for a single user.
  • Trust in the operator: Today's Muse Secure VM isolates each user from other users and external attackers, but the virtual machine memory is not cryptographically inaccessible to Meta itself. To address this gap, Meta has promised a Muse Confidential VM later in 2026 built on hardware-enforced confidential computing that will encrypt the entire virtual machine, including stored credentials and conversation history, with keys held exclusively on the user's device.
  • Platform permission and blocking: External websites ultimately decide whether to admit automated browser agents. On September 21, 2026, Amazon blocked Muse from accessing its online store, stating that Meta had not coordinated with Amazon ahead of launch and that the agent failed to identify itself transparently while browsing.
  • Scale of agent traffic: By late September 2026, bot-management provider HUMAN Security attributed 72% of all daily AI agent web requests across its monitored network to Muse alone. Around October 12, bot-detection firm DataDome observed that Meta began cryptographically signing a portion of Muse's outbound HTTP requests using the emerging Web Bot Auth standard so publishers and merchants can verify authentic Muse traffic.
  • Compute capacity limits: Running persistent cloud virtual machines and multimodal reasoning loops is computationally expensive, so the free tier carries a weekly usage cap of roughly 100 million tokens according to Mark Zuckerberg, while frontier models like Muse Wonder and Muse Wonder Full are gated behind paid subscriptions.
  • Behavioral fallibility: Meta warns directly inside the interface that "Muse may take unexpected actions," and early reviewers advise monitoring the activity feed closely, especially for email replies, retail purchases, and restaurant or travel reservations where errors are difficult to reverse.

These physical and architectural constraints directly shape how Muse should be configured in practice. Granting narrow connector permissions, choosing read-only access wherever possible, and keeping Sentinel set to "Always ask" for outbound actions dramatically reduce the blast radius of any single model hallucination or prompt injection attack, while the immutable audit trail lets users verify exactly what the agent did. The same trade-offs apply to every autonomous agent that acts with live user credentials, echoing foundational cybersecurity principles: grant software the minimum privilege required to complete the job, isolate untrusted execution environments, and keep a human in the loop for any irreversible commitment.

Why This Matters

Muse represents Meta's primary bid to turn its massive infrastructure investment in artificial intelligence into a daily utility that consumers and small businesses will pay for directly. At the same time, it serves as a high-stakes market test of whether a corporation whose business model was built on behavioral advertising can earn enough public trust to manage people's private communications, finances, and digital errands. Several factors make the rollout an inflection point for the broader AI industry:

  • Distribution at Meta scale: Muse climbed to the top of Apple's free App Store chart roughly 10 days after its US launch, outpacing ChatGPT's early mobile adoption curve according to app-intelligence estimates, and by October 21, 2026, Meta reportedly counted more than 10 million completed account setups across the United States and Canada.
  • A diversified revenue model: Monthly subscriptions ($20 Power and $100 Maximum tiers), Muse Code developer seats, and merchant transaction fees give Meta a direct software revenue stream that does not depend on advertising, while Muse for Small Business extends those same agentic workflows to local merchants and service providers.
  • Agentic commerce disruption: Deep integrations with Stripe, Shopify, PayPal, and major retail and travel brands including Best Buy, Gap, Sephora, Walmart, Wayfair, Expedia, and Instacart turn the agent into an intermediary storefront. Shares of online travel agencies Airbnb, Expedia, and Tripadvisor fell roughly 5% during the week of Meta Connect, reflecting investor concern that personal agents will disintermediate traditional booking sites and capture customer loyalty.
  • A reference security architecture: Analysts at DeepLearning.AI's The Batch noted that Meta engineered operating-system-level containment controls designed to hold even when the underlying language model is tricked, putting into production dual-agent separation patterns that security researchers had long advocated for systems combining private credentials, untrusted web content, and external network access.
  • Toward personal superintelligence: Meta executives frame Muse as the foundational software layer for the consumer "personal superintelligence" vision Mark Zuckerberg has outlined, spanning smartphones, desktop computers, talking Realtime Avatars, smart glasses, and dedicated wearable hardware like Muse Charm.

In the consumer and prosumer market, Muse competes directly with xAI's Grok Bot, OpenAI's dots, Google's Gemini Spark, Anthropic's Claude Cowork and Claude Duet, and Perplexity Computer, while Muse Code competes in the developer terminal against Claude Code and Codex. Side-by-side feature breakdowns are maintained in the Ready-to-Use AI Agent Comparison and AI Coding Agent Comparison, and independent evaluation suites such as the Assistant Benchmark track Muse alongside its rivals on multi-step real-world tasks, though rapid weekly updates mean rankings shift frequently. Because each of these competing agents makes different architectural choices about cloud versus local execution, native service connectors, and confirmation frequency, the most effective tool for any individual depends heavily on the software ecosystem they already inhabit.

Building Blocks

Component What It Does Notes
The brain: Muse model family Muse Spark, a natively multimodal reasoning model from Meta Superintelligence Labs, powers the agent; Muse Spark 1.3 shipped on September 2, 2026, and the frontier Muse Wonder family followed on October 12. The broader family includes Muse Image, Muse Video, the open-weight Muse Glimmer (30B) and Muse Spark Open (816B MoE) models, and Muse Wonder Full for coding.
Muse Code A terminal-based coding agent for software engineers that reads repositories, edits multi-file code, runs shell commands, and executes test suites. Debuted with Muse Spark 1.2 in August 2026; offers exclusive access to the unpruned Muse Wonder Full model for paid subscribers.
Muse Secure VM A dedicated, isolated cloud virtual machine with its own built-in browser that houses the user's agent, files, and session state. Swaps surrogate tokens for real credentials at the network boundary so the model never sees passwords. A hardware-encrypted Confidential VM is slated for late 2026.
Sentinel A separate system-level permission agent that approves, denies, or escalates every connector action and outbound network request. Approval prompts render as unspoofable system dialogs in the Muse app rather than as chat messages.
Connectors and integrations Links to consumer and developer services such as Google Workspace, Spotify, OpenTable, Instacart, Notion, and GitHub, and can configure connections to public REST APIs. The macOS desktop app also interacts with local Apple Messages, Calendar, and Notes, with broader macOS application control announced at Meta Connect.
Agentic Wallet Stores payment methods through Link by Stripe and executes checkouts with single-use virtual cards, alongside native Shop Pay and PayPal checkout support. Sentinel pauses execution and requires explicit user approval before every purchase.
Proactive goals and Ideas Tracks ongoing multi-week goals alongside one-off errands, compacts long-term memory in the background, and proposes timely actions in an Ideas tab. Users can inspect, edit, or delete remembered facts at any time.
Personalization and Realtime Avatar Lets users rename their agent and customize its visual avatar and communication tone. Meta's Realtime Avatar model, demonstrated with the mascot "Jolly," enables low-latency video-style conversations with a lifelike companion.
Agent email address Assigns a dedicated email address to each user's agent so people can forward receipts, newsletters, or threads directly into its queue. Announced at Meta Connect on September 23, 2026.
User and privacy controls Granular settings govern per-connector read/write access, Sentinel confirmation rules, and a complete audit log of agent actions. Data controls allow users to opt out of having their interactions used to train Meta's AI models, though training is enabled by default.
Consumer and Pro plans A free tier with a ~100M weekly token cap for everyday tasks, plus Power ($20/month) and Maximum ($100/month) subscriptions for heavy compute and priority features. Frontier reasoning upgrades like Muse Wonder roll out to paid subscribers first, while Muse Code has dedicated access to Muse Wonder Full.
Muse for Small Business Extends the Muse app with commercial skills and integrations for managing storefronts, customer outreach, accounting summaries, and ad campaigns. Launched September 29, 2026, with 15 core connectors including Shopify, QuickBooks, Stripe, Canva, Klaviyo, Slack, Zoom, Instagram, and Meta Ads.

Getting Started

Setting up Muse takes only a few minutes, and the most consequential decisions involve defining security boundaries and connector permissions rather than learning complex command syntax. Technology reviewers who stress-tested Muse during its launch window consistently recommend starting with strict confirmation rules and low-stakes organizational tasks, then gradually expanding permissions as you observe how the agent interprets instructions. The onboarding workflow below synthesizes Meta's official documentation and practical setup guides from Engadget and early testers:

  1. Check eligibility: You must be 18 or older and located in a supported region (the United States, Canada, Japan, or Singapore as of late October 2026). You need an active Meta account, though not a Facebook or Instagram profile, and launch documentation noted that a valid payment card must be placed on file during registration even to use the free tier.
  2. Get the app: Download the standalone Muse app for iOS or Android, open muse.ai in a desktop or mobile web browser, or install the native macOS application. In supported regions, you can also initiate tasks by messaging Muse directly inside WhatsApp.
  3. Sign in: Authenticating with your phone number on the web or mobile app can automatically discover and link your existing Meta account credentials.
  4. Personalize your agent: During onboarding, Muse prompts you to give the agent a custom name and select an avatar and conversational style. While the visual avatar is largely cosmetic, choosing a distinct name makes it much easier to address the agent naturally when copying it into email threads or messaging it on WhatsApp.
  5. Opt out of model training if preferred: Navigate to Settings, open the Data controls menu, and deselect "Help improve our AI models" if you do not want your prompts, files, and agent trajectories used to train future Meta models.
  6. Configure Sentinel permissions: While the default permission tier is "Ask for some actions," switching the global setting to "Always ask" during your first week ensures Sentinel will halt and display a system confirmation dialog before any external message, calendar change, or transaction occurs.
  7. Connect external services selectively: Open the Connectors tab to link accounts such as Gmail, Google Calendar, Spotify, OpenTable, or GitHub, selecting read-only scopes wherever the option is offered. Security-conscious users often create a dedicated secondary email account or forwarding rule specifically for their agent, while merchants can activate commercial integrations like Shopify, Stripe, and QuickBooks through Muse for Small Business.
  8. Fund the wallet only when needed: When you are ready to let Muse make purchases, add a payment method in the Wallet section. Link by Stripe tokenizes the card so the Muse runtime never sees your raw account number, and Sentinel requires manual approval before any checkout completes.
  9. Hand off an initial low-stakes task: Describe the final outcome you want, specify hard constraints such as dates, budgets, or preferred brands, and state explicitly where the agent should pause for review. Begin with research, summarization, or itinerary drafting before delegating tasks that send external communications or spend money.
  10. Monitor the activity feed and audit trail: Watch the live browser view while Muse navigates complex sites, check the Ideas tab for proactive follow-ups, review the chronological audit log after tasks finish, and instruct Muse to "forget" any sensitive details you do not want retained in long-term memory.
  11. Select an appropriate compute tier: Remain on the free plan while evaluating how Muse fits your daily routine, then upgrade to Power ($20/month), Maximum ($100/month), or a Muse Code developer seat if you regularly exhaust the weekly token allowance or need access to Muse Wonder and Muse Wonder Full.

Establishing a few disciplined habits early makes the difference between a reliable digital assistant and a privacy or financial liability. Always keep mandatory human approval enabled for outbound email, financial checkouts, and binding reservations, where automated misunderstandings are hardest to undo. Users should also avoid generating public share links for conversations that contain personal identifiers, financial records, or proprietary business data, a risk underscored when Meta suspended its chat-sharing feature altogether in October 2026 after thousands of shared links were discovered in public DNS logs. Above all, treat every draft, schedule, or cart Muse assembles as a proposal that you remain personally responsible for verifying before execution.

Examples

Meta's product documentation, independent reviewers, and company executives have demonstrated a wide variety of multi-step workflows that Muse and Muse Code can execute across consumer, creative, and commercial domains. The examples below are grouped by category, and nearly all of them follow the same operational arc: the agent gathers context across multiple connected services or local files, formulates a structured action plan or draft, and pauses at a Sentinel checkpoint before committing any consequential change.

Everyday Life

  • Recipes and grocery logistics: Extract ingredients from a cooking reel saved on Instagram and turn them into a consolidated shopping cart on Instacart, or pull decades of recipes saved across Facebook, Instagram, and bookmarks into a single searchable, categorized cookbook.
  • Entertaining and events: Design a balanced dinner party menu, automatically cross-reference remembered dietary restrictions for each invited friend, check calendar availability, and draft personalized invitations.
  • Travel planning and monitoring: Research complex international itineraries, compare nonstop flight options across carriers, and run a background price tracker that alerts you only when fares drop below a target threshold. Mark Zuckerberg has also described using Muse to monitor and secure hard-to-get mountain climbing permits as soon as recreation portals release them.
  • Family administration: Compile a weekday morning family briefing by cross-referencing shared calendars and incoming emails, extract key dates from lengthy school PDF attachments, or format a complete family newsletter as a polished, printable PDF in a single turn.
  • Personalized city guides: Scan years of saved social media posts and map bookmarks to build a curated restaurant and activity guide organized by neighborhood for an upcoming trip.

Money and Errands

  • Lowering recurring bills: Research competing broadband, mobile, or insurance promotional rates and contact the current service provider, including via automated voice or chat channels, to negotiate a lower monthly rate.
  • Buying and selling goods: Research fair market values to help sell a used car for a higher price, draft and manage listings on Facebook Marketplace, monitor stock availability across retail stores, and complete approved checkouts using single-use Stripe virtual cards or Shop Pay.
  • Sourcing local services: Reach out to multiple local plumbers, electricians, or repair shops to compare rates and find the earliest available appointment window that fits your calendar.

Goals, Projects, and Software Engineering

  • Adaptive fitness coaching: Maintain a multi-month marathon or strength training plan that automatically recalculates weekly workouts whenever work travel or illness disrupts your calendar.
  • Complex hobby and hardware builds: Coordinate multi-week personal projects, such as the weekend baking experiments and custom robot builds Zuckerberg plans with Muse, keeping track of bill-of-materials spreadsheets, vendor lead times, and sequential assembly steps.
  • Skill acquisition: Transform an open-ended learning goal into a structured syllabus of practice sessions, reference materials, and periodic check-ins that Muse proactively resurfaces in the Ideas tab.
  • Autonomous software development (Muse Code): Point Muse Code at a local software repository from the terminal to diagnose a failing continuous-integration build, refactor a legacy module across dozens of files using Muse Wonder Full, write unit tests to verify the fix, and open a documented pull request on GitHub.

Small Business

Even before Meta released dedicated commercial tooling, Chief AI Officer Alexandr Wang noted that early adopters were already using the consumer Muse app to help run plumbing contractors, independent grocery stores, family farms, restaurants, and retail shops. With the launch of Muse for Small Business on September 29, 2026, business owners can link Shopify, QuickBooks, Stripe, Canva, Klaviyo, Slack, Zoom, Instagram professional analytics, Facebook Pages, and Meta ad accounts inside the same workspace. That connectivity allows the agent to analyze weekly sales trends, identify slow-moving inventory, design promotional graphics in Canva, draft targeted customer emails in Klaviyo, prepare bookkeeping reconciliation summaries, and propose budget shifts across Meta ad campaigns, while holding every post, message, and expenditure in draft until the business owner approves it.

Sample Requests

The table below illustrates representative natural-language requests documented by Meta and independent reviewers, pairing each prompt with the background actions Muse performs and the specific checkpoint where Sentinel requires human approval. Because Muse interprets ordinary conversational intent rather than rigid command syntax, clearly stating the desired outcome, budget or timing constraints, and expected approval boundaries matters far more than phrasing.

Request What Muse Does Where You Approve
"Turn the recipe reels I saved on Instagram into a grocery list for this week." Reads saved Instagram posts, extracts and deduplicates ingredients, and assembles an online cart through a grocery connector such as Instacart. Before placing the delivery order or charging the wallet
"Plan a dinner for eight on Saturday and remember that Sam is vegetarian." Proposes a multi-course menu that respects stored dietary preferences, builds a prep timeline, and drafts calendar invitations. Before sending invitations to guests
"See whether I can lower my monthly internet bill." Benchmarks competing local broadband offers and contacts the provider's retention department to negotiate a rate reduction. Before agreeing to any contract renewal or plan change
"Research a March trip to Tokyo with nonstop flights and watch the prices." Compares nonstop flight schedules and hotel rates, sets up an asynchronous background price monitor, and notifies you when fares drop. Before booking flights or rooms
"Every weekday morning, brief me on my calendar and anything urgent from my kids' school." Scans connected calendars and inbox threads, parses PDF school newsletters, and compiles a concise morning summary. Before sending any reply to teachers or organizers
"Review last month's Shopify sales and draft a promotion for our slowest-selling items." Queries Shopify inventory and order metrics via Muse for Small Business, generates visual assets via Canva, and drafts a Klaviyo email campaign and Facebook post. Before any discount is activated or campaign is published
"In this repo, find why the payment webhook test is failing on main and fix it." (via Muse Code) Reads test logs, traces the bug across source files using Muse Wonder Full, edits the implementation, and re-runs the local test suite until it passes. Before executing destructive shell commands or pushing commits to remote

Known Limitations

Despite its rapid consumer adoption and layered security model, Muse faces significant technical, operational, and privacy limitations that prospective users must weigh before granting it broad account access:

  • Geographic and age restrictions: As of late October 2026, Muse is officially available only to adults aged 18 and older in the United States, Canada, Japan, and Singapore, and Meta has not published a rollout schedule for the European Union, the United Kingdom, or other major regions where regulatory scrutiny of autonomous data processing is strict.
  • Upfront payment card requirement: Launch reporting documented that users are required to place a valid payment card on file during onboarding even if they only intend to use the free tier, creating friction for privacy-conscious consumers.
  • Model fallibility and injection exposure: Meta warns in the product interface that Muse may take unexpected actions, and its system card acknowledges that Muse Spark 1.3 remains vulnerable to adaptive jailbreaks and indirect prompt injection when browsing untrusted websites or parsing malicious emails.
  • Operator access to cloud VMs: Until the hardware-encrypted Muse Confidential VM ships later in 2026, Meta's infrastructure engineers technically retain the ability to access the Muse Secure VM environment that holds a user's active session state and workspace data.
  • Default model training and user profiling: User interactions and agent trajectories are used to train Meta's AI models by default unless the user manually opts out in Settings. Furthermore, an October 2026 investigation by TIME revealed that Muse's hidden system instructions direct the agent to build a detailed profile of the user and their personal contacts, share anonymized behavioral insights with Meta to improve products, and retain original chat messages on Meta servers even after a user explicitly instructs the agent to "forget" a piece of information.
  • Public exposure of shared chats: On October 20, 2026, Meta abruptly suspended the ability to share Muse conversations after cybersecurity firm Claroc discovered roughly 6,800 shared Muse chat URLs exposed through public DNS records, many of which contained sensitive personal and commercial records including loan applications, internal revenue reviews, and private correspondence.
  • Merchant and website blocking: Major platforms can block agentic traffic at the network edge; Amazon blocked Muse on September 21, 2026, over transparency and notification concerns, and while Meta began testing cryptographic Web Bot Auth signatures in mid-October, acceptance across the wider web remains inconsistent.
  • Undisclosed human-in-the-loop testing: An investigation by 404 Media in September 2026 revealed that during pre-launch beta testing, certain outbound phone calls presented to users as automated AI negotiations were actually completed behind the scenes by human call-center contractors; Meta responded that human-in-the-loop evaluation is standard industry practice during testing and pledged clear disclosures for public voice features.
  • Closed connector review and MCP absence: While enterprise developers can build custom integrations, third-party listings in Meta's official connector directory must pass a multi-stage security review, and the consumer Muse app does not offer an open, general-purpose Model Context Protocol (MCP) configuration menu for connecting arbitrary local tools.
  • Unannounced hardware pricing and specs: Although Meta previewed the palm-sized Muse Charm companion device at Meta Connect for a December 2026 release, the company has not disclosed its retail price, battery life, or complete hardware specifications.

History

Recent Developments

The development and rapid commercial expansion of the Muse ecosystem unfolded over a compressed thirteen-month timeline at Meta:

  • September 2025: An engineering group led by Nat Friedman, who joined Meta to oversee AI product development alongside Chief AI Officer Alexandr Wang at Meta Superintelligence Labs (MSL), began building a consumer and developer agent platform under the internal codename "Hatch," drawing architectural inspiration from self-hosted agent frameworks like OpenClaw.
  • April 8, 2026: Meta Superintelligence Labs officially introduced Muse Spark, a natively multimodal reasoning model and the foundational architecture of the Muse model family. An internal target to launch the consumer Hatch agent in April was postponed so security teams could harden the runtime against prompt injection and credential leaks.
  • May 2026: The Information reported details on the internal Hatch prototype, noting that Meta engineers had initially prototyped the agent harness using Anthropic's Claude Opus 4.6 and Claude Sonnet 4.6 models while waiting for Meta's proprietary Muse Spark checkpoints to reach parity on complex tool use.
  • August 2026: Meta began expanding the model lineup rapidly, releasing Muse Spark 1.2 on August 5 alongside the debut of Muse Code, Meta's terminal-based coding agent for software developers, followed by Muse Glimmer, a compact 30-billion-parameter open-weight model under the permissive Apache 2.0 license, on August 10.
  • September 2, 2026: Meta shipped Muse Spark 1.3, featuring upgraded visual chain-of-thought, parallel sub-agent "Contemplating" mode, and stronger adversarial robustness to power the upcoming consumer launch.
  • September 8, 2026: Meta officially launched Muse in the United States across iOS, Android, muse.ai, and WhatsApp, introducing the Muse Secure VM, the Sentinel permission agent, Link by Stripe checkout, the free tier, the Power ($20/month) and Maximum ($100/month) subscriptions, and a security bug bounty offering rewards up to $300,000.
  • September 17 to 18, 2026: Meta released a native macOS desktop application capable of interacting with local Apple apps, expanded Muse availability to Canada, and reached the number-one position on Apple's US free App Store chart.
  • September 21 to 22, 2026: E-commerce reception diverged sharply as Amazon blocked Muse from scraping or transacting on its marketplace on September 21, while on the very same day Shopify announced native agentic checkout via Shop Pay across all Shopify merchants, followed by PayPal enabling merchant checkout for Muse on September 22.
  • September 23, 2026: At its annual Meta Connect conference, Meta unveiled a low-latency Realtime Avatar model (demonstrated with the animated character "Jolly"), deeper macOS desktop application control, dedicated agent email addresses, hands-free wake-word access on Meta AI glasses, expanded retail partnerships (including Best Buy, Walmart, Expedia, and Instacart), and Muse Charm, a pocket-sized hardware companion slated for December 2026.
  • September 29, 2026: One day after introducing the broader Meta Enterprise Platform, Meta launched Muse for Small Business, equipping the agent with 15 commercial connectors including Shopify, QuickBooks, Stripe, Canva, Klaviyo, Slack, and Meta Ads.
  • October 6, 2026: TIME published an investigative report examining Muse's internal system prompts, highlighting how the agent maps user social graphs and retains raw conversation logs even after users ask it to forget specific facts.
  • October 12, 2026: Meta unveiled Muse Wonder and Muse Wonder Full, its most capable frontier agent models to date; standard Muse Wonder began rolling out to consumer Power and Maximum subscribers, while the full-scale Muse Wonder Full model was made exclusive to paid Muse Code users. Around the same date, security researchers at DataDome observed Meta testing cryptographic Web Bot Auth signatures on outbound Muse browser traffic.
  • October 15, 2026: Marking a high-profile return to frontier open-weight releases following the LLaMA series, Meta released Muse Spark Open, an 816-billion-parameter (42-billion active) mixture-of-experts model with a 1-million-token context window under a custom commercial license featuring a $20 million annual revenue threshold for hosted cloud providers.
  • October 20 to 21, 2026: Meta suspended the public conversation-sharing feature in Muse on October 20 after Claroc disclosed that roughly 6,800 shared chat links had leaked into public DNS records; the following day, internal figures cited in press reports indicated Muse had surpassed 10 million account setups in the US and Canada.
  • October 23 to 27, 2026: Meta accelerated international expansion in Asia, launching Muse in Japan on October 23 with regional travel partners like Klook (where it hit number one on Japan's App Store within 24 hours according to Alexandr Wang) and launching in Singapore on October 27 with localized merchant connectors.

Further Resources

Additional Viewing

Load video
YouTube

Introducing Muse: your personal AI agent is Meta's official launch video published on September 8, 2026, accompanying the company's newsroom announcement. It demonstrates how Muse moves beyond conversational answers to execute multi-step errands across web browsers and connected apps, serving as a concise visual overview for the Introduction/Overview section. Watch on YouTube

Load video
YouTube

Mark Zuckerberg's Plan for a Personal AI That Works For You | Meet Muse is an in-depth launch-day interview hosted by the Tiff In Tech channel in which Mark Zuckerberg explains the architectural and philosophical foundations of personal superintelligence. He details the 100-million-token weekly free allowance, the Muse Secure VM and upcoming Confidential VM, one-time Stripe virtual cards, the Sentinel permission agent, background memory compaction, and his personal use of Muse for hobby builds and mountain climbing permits, directly supporting the Mechanisms and Examples sections. Watch on YouTube

Load video
YouTube

Make money, cut costs & run your life with Meta's Muse is a practical hands-on review from The Next New Thing, hosted by Andrew Warner and published on September 28, 2026. It walks through live demonstrations of Muse negotiating recurring bills by phone, tracking retail prices, checking out with virtual cards, managing Facebook Marketplace listings, parsing school PDFs, and compiling daily family briefings while highlighting where Sentinel approval prompts slow execution, mapping directly to the Examples and Known Limitations sections. Watch on YouTube

Load video
YouTube

"Muse gets AI agent UX right" is a 30-plus-minute episode of How I AI published on September 16, 2026, in which host Claire Vo stress-tests Muse across real-world personal and professional workflows. She evaluates the onboarding sequence, calendar management, long-term goal tracking, one-shot PDF newsletter generation, live browser shopping, the animated avatar, and the Sentinel permission model while comparing the user experience against Anthropic's Claude and OpenAI's Codex tools, providing an essential companion to the Getting Started, Mechanisms, and Muse Code sections. Watch on YouTube

Load video
YouTube

Muse Spark Tested: Meta's Comeback Model That Rebuilt Everything From Scratch is a hands-on review by Fahd Mirza, published on April 8, 2026, the day Muse Spark launched, and running a little over 16 minutes. He introduces the model as a multimodal reasoning system with tool use, visual chain-of-thought, and multi-agent orchestration, then puts it through a series of practical prompts. It is a useful first look at the original release described in the Muse Spark and How Muse Spark Works sections. Watch on YouTube

Load video
YouTube

Meta Muse Spark 1.3 Review: A Huge Leap Forward is a roughly 20-minute review by Johanna Hendrix, published on September 2, 2026, the day Muse Spark 1.3 shipped. She compares it with Gemini 3.8 Flash and other current models on video understanding, customer-support replies, browser game coding, animations, and writing quality, and concludes that 1.3 is a large step up from 1.2. Watch it alongside the Versions and Benchmarks section, which explains why independent benchmark scores and hands-on results do not always agree. Watch on YouTube

Load video
YouTube

Meta Connect 2026: Opening Keynote is the official recording of Meta's September 23, 2026 keynote, published by the Meta Developers channel. It sets out Meta's roadmap for AI, AI glasses, and VR, including the new Ray-Ban Meta and Meta Glasses lineup, Meta Ray-Ban Display updates, Meta VR Glasses, and Muse on wearables. It is the primary source for the Connect announcements in the Meta AI on AI Glasses section and the Muse announcements later on this page. Watch on YouTube

Load video
YouTube

Introducing My Computer: When Manus Meets Your Desktop is Manus AI's official launch video for My Computer, published on March 16, 2026, while Manus was owned by Meta. It shows the agent sorting large numbers of disorganized files, building desktop apps with local development tools, using an idle GPU to train models or run language models, and taking instructions from a phone while the home computer does the work. It is the best visual introduction to the Manus My Computer section and its approval-based safety model. Watch on YouTube

Load video
YouTube

Introducing Manus 2.0 is the official announcement video published by Manus AI on September 28, 2026, four weeks after Manus resumed independent operations. It introduces Cascade, Manus Studio, the Video Editor, Game Dev, Cloud Computer, Remote Control, Automations, and Cue, the new personal-agent app. Its short format makes it a useful orientation before the Manus offerings sections. Watch on YouTube

Load video
YouTube

Manus Demo Walkthrough 2026 is an official feature tour presented by Alan Chan of Manus AI and published on January 12, 2026. It demonstrates slide generation, Cloud Browser and Browser Operator, Wide Research, Mail Manus, web development, scheduled tasks, spreadsheets and data visualization, image and video generation, connectors, and projects. It directly supports the Manus Mechanisms and Cloud Computer, Browser, Wide Research, and Mail sections. Watch on YouTube

Load video
YouTube

Learn with Manus Webinar: Build, Publish, and Manage Your Web & Mobile Projects is the recording of an official Manus webinar held on April 23, 2026, and posted on April 28. It follows the lifecycle of building and publishing websites and mobile apps, covering databases, user authentication, analytics, custom domains, payments, embedded AI tools, app store publishing, and the desktop app. It provides visual context for the Manus Studio and Creation Environments section. Watch on YouTube

Load video
YouTube

How to use Manus for Brands: Webinar Recording is an official Manus for SMB webinar published on August 5, 2026, co-hosted by Yee Su, a serial founder with e-commerce and agency experience. It shows how to connect Manus to a Shopify store and works through four use cases: building a storefront, managing a product catalog, turning analytics into growth opportunities, and automating work with Brand Skills and scheduled tasks. It is particularly relevant to the Automations, Scheduled Tasks, and Connectors and Manus for Enterprise and Small Business sections. Watch on YouTube

Load video
YouTube

How to Scale Your CPG Brand with AI is an official Manus for SMB webinar recording published on July 20, 2026, featuring Slater Caskey, chief executive of the consumer packaged goods brand Claros Farm. He shows how he uses Manus as an AI chief of staff for procurement, inventory forecasting, search engine optimization, sales dashboards, and what-if purchasing models. It complements the small-business discussion and its caution that polished dashboards still depend on correct underlying data. Watch on YouTube

References

```

Retrieved from "https://primo.ai/w/index.php?title=Meta_AI&oldid=53497"