AgentComparison
Meta · product · personal assistant

Muse

Meta's personal agent for reminders, actions and file work.

Group: Personal agents you message

Meta documents Muse as a personal agent for actions, reminders, ongoing conversations and file work.

Documentation researched, not testedChecked 28 October 2026Vendor documentation, not a test

The decision in plain English

Review actions and define boundaries: Meta warns that Muse may be inaccurate or take unexpected actions. Pricing is now documented (free weekly limit; Power $20/mo; Maximum $100/mo) but regional availability is not - check the app shows Muse as available before recommending. Muse is not in our text-message assistants table under the published inclusion rule: Meta documents conversations in its own app and web chat only, with no messaging-service channel (iMessage, SMS, WhatsApp, Telegram, Slack, email or phone).

Best fit

Personal reminders, ongoing tasks and artifact work.

Poor fit

Unmonitored actions where errors would be costly.

Documented capabilities

Vendor statements, not verified task outcomes.

  • Reminders and ongoing conversation context Source · checked 2026-10-23
  • PDFs, spreadsheets and other artifacts Source · checked 2026-10-23
  • App and web access Source · checked 2026-10-23
  • Connectors: Meta and third-party app connections (Gmail/calendar-style read and write), retrieve-only mode per connector, custom connectors via chat, credentials in Secure Credentials Store. Facebook, Instagram and Threads auto-connect via Accounts Center. Source · checked 2026-10-28
  • Built-in web browser: booking appointments, purchases, online forms and price-watching, with a live viewing window, take-control pause and stop. Browsing credentials stored in the Secure Credentials Store. Source · checked 2026-10-28
  • Approval system: per-connector read/write gates, allow once/task/site/always/deny decisions, activity log, scheduled-task approvals and allowed-website list. Source · checked 2026-10-28
  • Memory and identity files: Memories plus editable Memory.md, Soul.md and Identity.md; a forget skill erases topics; full reset permanently deletes chats, files and active tasks. Goals tab tracks personal goals with check-ins. Source · checked 2026-10-28

Permissions and human control

  • Runs from the Muse app or web under a Meta Account (Facebook/Instagram/email/phone signup paths).
  • Connectors for Meta and third-party apps (email and calendar examples; Facebook, Instagram and Threads auto-connect via Accounts Center). Per-connector read vs write can be set to retrieve-only. Custom connectors can be created by asking Muse; Meta does not review them.
  • Granular approvals: per connector 'Ask for some actions' (write and important reads) or 'Always ask'; web-access defaults; approval options are Allow once / for this task / for this site / Always allow / Deny. Meta says Muse will not take many important actions, like sending an email, without approval by default.
  • Built-in browser for bookings, purchases, form-filling and price-watching; user can watch live, take control (pause) or stop the task. Credentials sit in a Secure Credentials Store the model does not see.
  • Activity log, scheduled-task and artifact approvals, and an allowed-websites list are user-manageable in Settings.
  • Can send emails, texts and artifacts to others; Meta suggests asking it to attribute AI-generated content to the agent. Some actions (like a sent email) cannot be reversed.
  • Creates and edits artifacts (PDFs, spreadsheets, documents, images, podcasts) saved to a Library workspace.

Documented controls do not prove that an agent always follows them.

Limits and open questions

  • Meta warns of inaccurate responses or unexpected actions and tells users to review activity regularly.
  • Some actions, like a sent email, cannot be reversed.
  • AI-training opt-out is a manual setting that starts ON.
  • Custom connectors are not reviewed by Meta.
  • Limited testing: no supported-country list published; subscription management is locked to the original platform, device type and store account.
  • Free usage limit size and token-to-task equivalence are unpublished. Reliability and task success were not tested.

Privacy and security

Meta's Muse help pages (re-read 28 October 2026) state: each user's Muse runs in an isolated cloud VM, with a Confidential VM option planned that Meta says is designed to keep information private even from Meta. Muse does not share conversations or VM data with Meta ad systems. Credentials live in a Secure Credentials Store and are used only after checks that the service, authorization and permission type match; the AI model does not see passwords. Meta documents prompt-injection mitigations (model training, automated detection, permission checks operating separately from the model). AI-training on interactions is ON by default (Settings > Data controls); turning it off stops use for model improvement and applies to previous interactions; when on, Meta says it removes certain PII categories and disassociates interactions from the account. Connector data exchange is scoped to what the task needs (email search does not download the whole inbox unless told); Google Workspace API use adheres to the Google User Data Policy Limited Use requirements. Custom connectors are not reviewed by Meta - grant access with caution.

Feature-by-feature: what is documented and what is not

Official sources checked 28 October 2026. A documented feature is a vendor statement, not proof of success.

FeatureEvidence statusDetail and limit
PricingFree with usage limitMeta help: Muse is free with a usage limit; hitting the limit means upgrading to a paid subscription or waiting for the refresh. No paid-tier price published on the help pages checked. Source
Account modelMeta AccountApp or web access via Meta Account; alternative signup via Facebook, Instagram, email or phone number. Source
Safety postureUser-monitoredMeta's own documentation: Muse 'may be inaccurate or take unexpected actions - monitor its actions carefully and intervene as needed.' Boundary-setting is left to user instructions. Source
MemoryOn by designRemembers conversation context across main chat and side chats; name, personality and preferences are customizable by instruction. Source

How to start with Muse narrowly

  1. Sign in with a Meta Account (or create one via Facebook/Instagram/email/phone) and start on the Chat tab.
  2. State boundaries explicitly in your first instructions (what it may search, send or change) - Meta's own guidance relies on user-set limits.
  3. Review its first task outputs and enable the AI-attribution note before letting it email or text anyone on your behalf.

Official product/setup source. These are documentation-based steps, not a hands-on setup test.

Common questions about Muse

What is Muse for?

Personal reminders, ongoing tasks and artifact work. The vendor describes it as: Meta documents Muse as a personal agent for actions, reminders, ongoing conversations and file work.

Who should look elsewhere?

Unmonitored actions where errors would be costly.

How much does Muse cost?

Meta's subscription help pages (read 28 October 2026): free with a weekly usage limit that refreshes; Power $20/month (500M Muse tokens/week); Maximum $100/month (3B tokens/week). Subscribe at muse.ai or in the mobile app; manage and cancel on the same platform, device type and store account used to subscribe. Meta says Muse and subscriptions are in limited testing and not available in all locations.

What can it access and what needs my approval?

Runs from the Muse app or web under a Meta Account (Facebook/Instagram/email/phone signup paths). Connectors for Meta and third-party apps (email and calendar examples; Facebook, Instagram and Threads auto-connect via Accounts Center). Per-connector read vs write can be set to retrieve-only. Custom connectors can be created by asking Muse; Meta does not review them. Granular approvals: per connector 'Ask for some actions' (write and important reads) or 'Always ask'; web-access defaults; approval options are Allow once / for this task / for this site / Always allow / Deny. Meta says Muse will not take many important actions, like sending an email, without approval by default. Built-in browser for bookings, purchases, form-filling and price-watching; user can watch live, take control (pause) or stop the task. Credentials sit in a Secure Credentials Store the model does not see. Activity log, scheduled-task and artifact approvals, and an allowed-websites list are user-manageable in Settings. Can send emails, texts and artifacts to others; Meta suggests asking it to attribute AI-generated content to the agent. Some actions (like a sent email) cannot be reversed. Creates and edits artifacts (PDFs, spreadsheets, documents, images, podcasts) saved to a Library workspace.

What happens to my data?

Meta's Muse help pages (re-read 28 October 2026) state: each user's Muse runs in an isolated cloud VM, with a Confidential VM option planned that Meta says is designed to keep information private even from Meta. Muse does not share conversations or VM data with Meta ad systems. Credentials live in a Secure Credentials Store and are used only after checks that the service, authorization and permission type match; the AI model does not see passwords. Meta documents prompt-injection mitigations (model training, automated detection, permission checks operating separately from the model). AI-training on interactions is ON by default (Settings > Data controls); turning it off stops use for model improvement and applies to previous interactions; when on, Meta says it removes certain PII categories and disassociates interactions from the account. Connector data exchange is scoped to what the task needs (email search does not download the whole inbox unless told); Google Workspace API use adheres to the Google User Data Policy Limited Use requirements. Custom connectors are not reviewed by Meta - grant access with caution.

Has AgentComparison tested it?

No. This record is built from the vendor's own documentation, checked 2026-10-28. Vendor claims are attributed, not confirmed.

How to evaluate Muse yourself

  1. Pick one low-risk task that matches its documented capabilities: reminders and ongoing conversation context.
  2. Write the expected output and what counts as failure before you start.
  3. Connect the minimum access needed, and check approval settings first.
  4. Run it twice and compare, then review every action it took.
  5. Check cost after any free allowance. Unknown price is not free.

Related

Documented use cases: Personal email, calendar and reminders.

Sources and history

Change history

Documentation overview expanded with current pricing/access sources. No hands-on results claimed.

Report a correction