Meta Muse: Features, Specs, Privacy and How It Compares

Published September 8, 2026. Source-based launch analysis; Kingy has not performed hands-on tests of the Muse personal agent.

Meta introduced Muse on September 8, 2026, as a personal AI agent for adults aged 18 and over, initially in the United States. People can communicate with it through its own app or WhatsApp and delegate everyday tasks, including email and travel arrangements. Associated Press launch coverage.

The appeal is easy to understand. Much of everyday administration consists of small jobs that cross several services: finding a date, reading a confirmation, comparing options, remembering a preference, and following up. An assistant that handles those steps competently could save time even without making a spectacular intellectual breakthrough.

Muse should be judged by how much supervision it requires to finish those jobs. A persuasive conversation is useful, but an incorrect booking or an overlooked deadline can erase the benefit. The relevant comparison is therefore with other products that execute tasks, maintain context, and return completed work.

The Muse name covers several different products

Today’s announcement concerns the personal agent. Meta has also used Muse for a family of models and related tools. Keeping those names straight prevents misleading specifications and comparisons.

For more detail on the underlying model, see Kingy’s Muse Spark 1.3 coverage. Readers considering a local deployment can also consult our Muse Glimmer guide.

This distinction matters when someone asks whether Muse is open source, how many parameters it has, or how much its API costs. Those questions may concern different layers: the consumer service, the underlying model, or a developer product.

The practical specification sheet

For a personal agent, useful specifications include execution environment, permissions, integrations, persistence, and output formats. A model’s context window alone says little about how well the finished service manages a week of tasks.

How the experience is designed to work

Meta’s designers describe a main conversation that accepts interruptions and multiple requests, with side chats for separate context. A Goals view tracks ongoing work; activity logs and editable memory files expose more of what the agent is doing. Its proactive notifications can be adjusted. How We Designed Muse.

Meta describes a Linux VM containing an isolated agent environment and separate security services. The agent can execute tools, while host-side services handle credentials and permissions. Browser tasks use a Chromium browser through a restricted interface. This structure aims to keep an agent processing untrusted content from controlling its own safeguards. Muse technical architecture.

The combination is significant because long-running tasks create a coordination problem. If a user changes a travel date while an agent researches hotels, the agent must attach that correction to the right job. If another task finishes at the same time, its update should remain understandable. These are product reliability problems as much as language-model problems.

Three illustrative tasks show what to evaluate:

• Manage a small event: Gather constraints, propose dates, build a supply list, and prepare invitations. Check whether the agent preserves the guest count, budget, and dietary requirements through revisions.
• Research a purchase: Compare three products using a fixed specification and total delivered cost. Check whether the result distinguishes available stock, shipping fees, and optional accessories.
• Maintain a project tracker: Turn an objective into milestones, update progress from new information, and surface overdue dependencies. Check whether completed work stays completed and changed requirements propagate correctly.

For each task, the important result is an accurate deliverable with a clear status. An agent should distinguish “prepared,” “submitted,” and “confirmed.” Those states have very different meanings to the person relying on it.

Meta’s store description advertises scheduling, subscription reviews, price tracking, fitness planning, and continued work after the app closes. These establish the intended use cases; they do not provide independent completion rates or evidence that every supported service works equally well. Muse from Meta listing.

Permissions and privacy deserve precise language

Axios reports that a separate Sentinel system decides whether proposed access to the internet and connected services is allowed, blocked, or referred to the user for approval. That separation is an architectural feature, not independent evidence that every harmful action will be caught. Axios launch coverage.

Meta says credentials are concealed from Muse, sensitive actions require approval, service access is revocable, and users can opt out of training. It also says conversations and VM data are excluded from its advertising systems. The stronger Confidential VM, with a user-held encryption key intended to prevent Meta access, is planned for later in 2026. It is a future feature. Meta’s privacy and security description.

The technical explanation adds two material qualifications. Launch-day protections restrict employee access through operational policies; they do not cryptographically prevent Meta from accessing data to operate, support, or secure the service. Sanitized interaction trajectories may be used for model training unless users opt out. Although VM data is excluded from Meta’s ad systems, activity Muse performs on external websites can indirectly influence advertising. Meta’s data policy explanation.

These statements concern different protections. Limiting what an agent can do, restricting how a provider uses information, and preventing the provider from reading information are separate properties. Describing all three as “completely private” would obscure the distinction readers need to assess.

The practical test is whether an approval makes the proposed action understandable: the recipient, information being shared, item being purchased, and applicable amount. Asking for confirmation helps only when the person can understand what they are authorizing.

The most relevant alternatives

The following is a comparison of documented capabilities and product emphasis. The suggested fit is editorial judgment, not a measured ranking.

Listed dollar amounts are US-dollar reference prices, not Canadian quotes, and do not normalize allowances or the work completed per subscription.

Gemini Spark is the closest comparison in product intent. Google’s documentation describes scheduled and event-triggered tasks, personal context, and local or remote browsing. It also publishes a limit of 15 simultaneous tasks. Its help page lists broader eligibility than its marketing overview, which still emphasizes Ultra; this comparison follows the detailed help page and treats actual account access as the final check. Gemini Spark help, Spark overview.

For ChatGPT, the current product name matters. OpenAI’s help center says the old ChatGPT agent mode is no longer available and directs users toward Work. Work is therefore the relevant current comparison; older references to Operator or agent mode describe earlier offerings. ChatGPT agent status.

Claude Cowork also extends beyond a local desktop session: Anthropic documents cloud tasks started from mobile, alongside desktop access to local folders and applications. Calling it a desktop-only assistant would miss a material part of the current comparison. Cowork documentation.

Choose evaluation tasks based on the work to be delegated, then give each eligible product the same inputs, constraints, and approval rules. A price comparison becomes useful after measuring how many satisfactory jobs fit within each allowance.

What the model performance claims do and do not establish

Meta reports that Spark 1.3 used approximately 20% fewer tool calls and 25% fewer tokens than Spark 1.2 in comparisons by its engineers. It also describes improvements in instruction following, multitasking, and recognizing when user help is needed. These are vendor-reported model improvements, not independently measured Muse app results. Spark 1.3 release.

Fewer tokens can reduce computational work, but do not directly tell a consumer how long a booking takes or how often a task succeeds. Browser delays, integrations, approval interruptions, retries, and the product’s chosen model settings all affect the experience. A coding evaluation also cannot establish that an agent is better at managing a family calendar.

A useful comparison would report task completion, factual errors, constraint violations, interventions, elapsed time, and cost or allowance consumed. It would also count false success reports, where the agent says a job is finished but the external service or deliverable shows otherwise.

The launch-day assessment

Muse presents a coherent proposition: delegate ongoing personal administration in a conversational setting, then review the consequential decisions. Its strongest potential advantage is reduced coordination effort. The value would come from remembering the right constraints, keeping track of work, and bringing a person back in at the right moment.

There is not enough evidence in the reviewed launch material to declare it the best personal agent or assign a defensible numerical rating. Nor does the announcement establish generally available access outside the US. The next useful evidence would be repeated, comparable tasks showing how often Muse completes work correctly and how much attention it needs along the way.

Disclosure: Exact Muse subscription prices, consumer quotas, and production reasoning settings were not verified as of September 8, 2026. Vendor descriptions establish intended capabilities; they do not establish independent task-completion results.

Featured image: Meta, from How We Designed Muse. Official promotional artwork, used to illustrate this launch report.