Meta Hatch AI Agent Leak Gives Way to Muse Launch and MCP Push

From leak to launch in seven weeks

In late August, The Information reported on internal Meta documents describing a consumer AI agent platform, codenamed "Hatch," that was due within weeks. The same documents described a new flagship model, codenamed "Watermelon," aimed at an October release.26 The headline number was the price. Meta had considered tiered subscriptions for Hatch, with the top tier at up to $199.99 a month for higher usage limits.36 Coverage tied both projects to Mark Zuckerberg's effort to earn money from Meta's huge AI spending and to build revenue that doesn't depend on advertising.42

Seven weeks later, a lot has happened, though not all of it matches the leak. Meta released a consumer agent called Muse on September 8. It is free up to a usage cap, with paid Power and Maximum plans at $20 and $100 a month.3132 Meta also opened a WhatsApp Business server built on the Model Context Protocol (MCP), an open standard for connecting AI agents to outside software.12 And it set up a new enterprise division to sell its models and agents to companies.17 Watermelon has not shipped. As the month runs down, prediction-market traders are betting it won't arrive by October 31.23

The main point is that Meta is no longer just running an agent experiment. It is building the pieces of a business: a consumer agent, protocol connections for developers, and an enterprise sales channel. How those pieces fit together, and what they will cost buyers, is still unclear.

What the Hatch reporting described

The early coverage agreed on the basics. Hatch was described as a consumer version of OpenClaw, an agent product it was meant to compete with. It was trained to work on sites including DoorDash, Etsy, Reddit, Yelp and Outlook. Prototypes showed a customizable dashboard of agent-built tools, such as a fitness tracker or a trip planner.26 Zuckerberg had told investors earlier in the year that he saw room for a more polished, easier version of OpenClaw. PYMNTS pointed to distribution as Meta's edge: OpenClaw has to be installed on a personal computer, while Hatch would live inside Instagram, which more than 2 billion people use every day.6 The reports also said a WhatsApp trial would let users connect third-party AI agents inside the messaging app.34

One outlet went further than the rest, reporting that Hatch was initially running on Anthropic's Claude Opus 4.6 and Sonnet 4.6 and would move to Watermelon over time.5 No other coverage repeated that claim, so it should be treated as unconfirmed. Later reports that Meta was shrinking its staff's internal use of Claude make it more interesting, though.

Outlets disagreed about how far along the project really was. Wired reported in September that employees had been testing Hatch on work devices for several weeks ahead of an expected public release. It described Hatch as Meta's most advanced AI experiment, an agent that can browse the web and run other applications on its own.44 Some staff were reluctant to connect it to personal email and calendars. They worried about privacy and about the agent breaking something important. Some of that distrust traced back to an earlier, now-paused project that tracked employee keystrokes to collect AI training data.44

Muse: Hatch under another name?

Meta has never publicly said that Muse is Hatch. Analysts noted early that Hatch and Watermelon were internal codenames and that Meta hadn't revealed what the consumer product would be called.1 One post-launch writeup treats Hatch as a separate effort from Muse.22 Still, what Meta shipped looks a lot like what was reported. Muse is a personal agent that runs errands and handles multi-step tasks. It is available on iOS, Android, the web and WhatsApp, US-only at launch, for adults.3532 My reading is that Muse is the public version of the Hatch work, at a much lower price than the leaked $199.99 ceiling. That is my inference, not something Meta has confirmed.

The pricing also differs from the leak in how it is built. Usage is measured in tokens, the units AI models use to count text. The free tier allows up to 100 million tokens a week, Power gives 500 million, and Maximum gives 3 billion.3134 Alexandr Wang, who leads Meta's AI work, said the plans are priced by usage rather than by feature.32 Meta has published the token limits, but it still hasn't said how many real tasks $20 or $100 buys. That makes it hard to compare Muse's price with other agents.31

The technical design is the most enterprise-relevant part of the launch. Each user gets a dedicated cloud machine, the Muse Secure VM, with its own browser. A separate agent called Sentinel has to approve anything Muse sends out to the internet.3438 Credentials are stored apart from the AI model, sensitive actions need human approval, and every action is logged.39 The limits matter, though. Meta's own security write-up says today's design doesn't stop Meta from accessing user data when it needs to run or secure the service. A Confidential VM, encrypted with a key only the user holds, is promised for later this year.36 During setup, Meta warns users that Muse "may take unexpected actions."38 Early numbers are decent: more than 2.5 million US downloads by September 23, and small-business connections to Shopify, QuickBooks, Stripe and Canva on September 29.31 Downloads are not paying users, however, and Meta hasn't said how many people have upgraded.31

MCP is how Meta reaches businesses

For enterprise readers, the more lasting change may be Meta's adoption of MCP, the Model Context Protocol. In mid-September, Meta opened a WhatsApp Business Tools MCP server. It lets outside coding agents such as Claude, Cursor, Codex or ChatGPT set up and manage WhatsApp Business messaging.12 Through it, an agent can create accounts, verify phone numbers, build and edit message templates, and flag problems like failed verification or payment errors.18 This adds to Meta's existing MCP servers for ads and app settings.12 Meta now sits alongside PayPal, Stripe, GitHub, Google and Microsoft as a platform that lets agents operate it directly.18

Meta's choice here stands out. It is building its own consumer agent and, at the same time, letting rival agents drive its business tools. The ads MCP server launched in open beta on April 29. It can read and change ad accounts, and changes take effect immediately, with no draft step.1519 Admins can set rules on what agents may change in budgets and product catalogs, but advertisers are effectively expected to check every live change themselves.15 Meta hasn't said what the server will cost after the beta ends.15 That is a real governance issue for any company letting autonomous agents touch money.

The pieces don't yet connect cleanly. Meta Business Agent, the company's agent for answering customers on WhatsApp, Messenger and Instagram, launched globally on June 3. It is charged at $2 per million tokens.14 One review notes that none of Meta's MCP servers can set up or run Business Agent itself.14

The enterprise platform and the Claude pullback

On September 28, Meta announced the Meta Enterprise Platform. Its first products are Muse Agent for internal task automation, Business Agent, Muse API for direct model access, and Muse Code, a coding agent priced against Claude Code and Codex. Former MongoDB CEO Chirantan Desai will run it.17 Zuckerberg called it the "next major pillar" of the business. Meta published no prices, launch dates, customers or regions.1728 Analysts pointed out that Meta has shut down past enterprise products, including Workplace, and said buyers will want proof of long-term commitment.17

Meta's internal behavior shows the same push to own the whole stack. Internal Claude Code use has reportedly fallen from about 60,000 employees to about 30,000. Staff are being steered toward Meta's own MetaCode and Muse Code. Some units restricted rival assistants over fears that a competitor's output could leak into Meta's training data. Even so, Meta reportedly spent more than $105 million on Claude Code in one 28-day period.

Watermelon and the reality check on agents

Watermelon is the part of the leak with the least to show. Reports said internal benchmarks claimed it matched GPT-5.5, using about ten times the computing power of its predecessor, Muse Spark.1 As of early October, there was no public launch, no launch event and no developer preview. Prediction markets show traders betting against a release before month's end.23

The wider evidence calls for caution. Wired reported that Meta's agent-driven layoffs were called off amid more software bugs linked to heavy AI reliance and slower-than-expected progress on agents.44 Separately, Meta has dropped AI usage and token counts from performance reviews.42 One analysis, citing internal posts reported by Reuters, connects unchecked agent activity to a 40% rise in major incidents.45

Taken together, the coverage shows Meta has a product, a protocol strategy and a sales team. It has not yet shown that autonomous agents are reliable enough to trust with real accounts and real spending. Until Meta publishes clear prices, independent test results and the promised encryption, companies should treat these tools as a fast-moving pilot, not finished infrastructure.