Meta Muse spaces: build your own app inside the agent, plus Wonder-trained models

Meta has added "spaces" to its Muse personal agent app: custom, app-like pages that the agent builds from a plain description, such as a gift tracker, a school dashboard or a trip planner. On the same day, 6 November 2026, the company said it is rolling out specialized models, trained with its frontier Muse Wonder model, to all Muse and Meta AI users. Together the two changes show where consumer AI agents are heading: less one-off chat, more persistent personal software that an agent writes and keeps up to date.

Key points

• Spaces are available "everywhere Muse is available" from 6 November, and are rolling out to all Muse users, according to Meta and The Verge.
• A user creates a space by describing it. With connected apps, the space refreshes in the background, Meta says.
• Meta is also shipping "specialized" models trained with learnings from Muse Wonder, with two features, Active Memory and Active Focus, to Muse and Meta AI.
• On Meta's own evaluation, the new models use critical context correctly in 90% of tasks, up from 76%, as reported by The New Stack. These are company figures, not independent tests.
• Meta says it is exploring sharing spaces with other people. That is not available yet.

What happened

Meta announced spaces in a newsroom post titled "Design the Muse App, Just for You". It describes a space as "a custom experience you create inside Muse by describing your idea". Its examples include a holiday gift tracker, a daily word game, a planning board for a DIY project, a wedding planner with an RSVP list and vendor invoices, an order tracker for a small business, and a trip planner with flights, hotel and a day-by-day itinerary.

Once created, a space can be changed by talking to Muse: adding what is missing, trying a new look or changing how it works. Users can place spaces in the sidebar or the main tabs, hide them, and schedule when they appear. Meta's example is to keep a weekly planner up front during the workweek and swap in weekend plans on Saturday. Inside a space you can check things off, add notes and edit directly, and the content is there when you come back.

Spaces can also connect to outside apps. Meta says that with connected apps "your spaces keep themselves up to date in the background", so new concert dates, price changes or savings progress appear without being asked for. The Verge lists Gmail, Google Calendar, OpenTable and Notion among the apps Muse works with.

TechCrunch's reading is more modest: it describes spaces as breakout conversations about a particular topic that live in the sidebar, while noting Meta calls them "custom experiences". The two descriptions are not contradictory, but they stress different things: Meta frames spaces as small apps the agent builds; TechCrunch focuses on the conversation each one keeps going. Mark Zuckerberg posted on X, per TechCrunch, that he had made "a BBQ game, an event tracker, and a home control dashboard".

The second announcement concerns the models underneath. The New Stack reports that Meta trained new models using Muse Wonder, its frontier model, and is shipping them to all Muse and Meta AI users, while subscribers still get Wonder itself for harder tasks. Meta has not named the new models. It calls them specialized versions of its existing models, trained for the behaviors a consumer agent needs.

Background

Muse, Meta's personal agent app, launched in September 2026 on the smaller Muse Spark model, with a free tier. The New Stack calls it "somewhat of a breakout hit". TechCrunch reports that on Meta's October earnings call Zuckerberg said the agents had 10 million users in the US and Canada, less than two months after they debuted. Meta launched Wonder and the more capable Wonder Full on 12 October, according to The New Stack. Wonder is available through the Meta Model API and in Muse Code, Meta's coding agent.

Spaces is the second Muse feature in a week. TechCrunch notes that Meta announced Muse Voice the day before, which lets users video chat with the agent and give spoken commands. The rivals are also named in the coverage: TechCrunch mentions OpenAI's dots and a Grokbot from SpaceXAI, and The New Stack contrasts Muse's free tier with OpenAI's Dots and Anthropic's Duet, which it says are only available on those vendors' highest pricing tiers. We covered Anthropic's Claude Duet launch separately.

The numbers

Meta published two internal measurements for the new models. They are the company's own evaluations and The New Stack did not see a baseline for the second.

Previous model76

New Muse models90

Share of tasks where critical context is used correctly, Meta internal evaluation (%)

	%
Previous model	76
New Muse models	90

Source: The New Stack, reporting Meta's announcement

For Active Focus, which keeps a model working when something goes wrong mid-task, Meta reports a 67% improvement in finishing routine tasks when the models hit unexpected problems. The New Stack notes Meta did not share the baseline, so the figure cannot be put in context. The API price for Wonder is $4 per million input tokens and $20 per million output tokens, per The New Stack.

What it means

For users, spaces turn the agent into a place to keep ongoing things. A conversation ends, but a tracker or planner that refreshes itself does not. That is a stronger reason to return to the app than a chat window, and it explains why Meta pairs spaces with memory features: Active Memory is described as training the models "to recognize what type of information is temporary versus permanent", which is what a long-lived planner needs.

For developers, the pattern is notable. An agent generating small interfaces on demand, wired to a user's accounts, overlaps with what many teams build by hand as dashboards and internal tools. It also raises the question of what these generated pages are allowed to do with connected accounts. Meta's post says users connect apps themselves, but it does not describe permissions, data handling or how generated spaces are checked, and we found no detail in the coverage we read.

For the model market, The New Stack's point is that Meta did not just shrink its large model. It trained the new models on consumer-agent behavior, such as deciding when to act alone and when to ask, and taking corrections mid-task, using a large model to train smaller ones. The article says Meta has not described its method. Serving lighter models to a free tier is also cheaper than serving Wonder itself to everyone, which is The New Stack's inference rather than Meta's statement.

What we don't know yet

• How spaces are built and sandboxed, and what permissions a space gets over connected apps. Meta's post does not say.
• Whether spaces are free for all users or limited by plan. Meta says spaces are available everywhere Muse is, without stating limits.
• The names, sizes and method of the new specialized models.
• Independent results for Active Memory and Active Focus. All figures so far are Meta's.
• When sharing and collaboration arrive. Meta says only that it is exploring them.

What to watch next

Watch for independent testing of the 90% and 67% claims, for Meta's documentation on how spaces use connected-app data, and for the launch of shared spaces. Competing agent products will likely add similar persistent views. Third-party developers may also want to know whether spaces will ever be exposed through the Meta Model API.

Sources

• Meta Newsroom: Design the Muse App, Just for You (primary source)
• TechCrunch: Meta introduces Spaces, a new way to collaborate with your Muse agent
• The Verge: Meta's Muse AI can create custom 'spaces'
• The New Stack: Meta's best AI model now trains the one most of its users get