Meta Muse review

Muse is worth considering if your day involves constant coordination between email, calendars, websites and other services. Its appeal is taking a task through research, contact, waiting and follow-up. Published reviews already describe useful results, alongside stale information, invented details and simple purchases that took longer than doing them manually.

The September 23–24, 2026 announcements add another layer of confusion: the original launch and the Connect updates are different events, while referral rewards and permission to join are different conditions. Neither a code nor a stage demonstration tells you everything about the product you can use today.

Method and fact-check date: September 25, 2026. This review combines official announcements, product explanations, policies and named journalists' experiences. We did not sign in, redeem a code or run Muse tasks. Reported hands-on results remain attributed to their authors.

What is Muse Agent?

Muse is Meta's personal AI agent, powered by Muse Spark. The product also includes a cloud computer, browser, memory, task management and connectors to other apps. Its practical usefulness depends on those pieces working together, as described in Meta's launch announcement.

You might ask it to compare restaurants, find a suitable time and work toward a reservation, returning when it needs your decision. According to the product design explanation, Muse can continue after you close the app, respond to schedules or relevant events, and keep goals in a Goals view. Activity logs show its work; memory files can be inspected and edited.

Our evaluation criterion is whether it reduces the work of gathering context, switching apps and chasing follow-ups, after accounting for the time needed to check its output. A fluent chat response alone cannot establish that.

What changed on September 23–24?

The original announcement is dated September 8. Connect presentations took place on September 23, followed by Meta's September 24 recap. The direction is broader access to the agent across devices and services, but release stages differ.

Alexandr Wang announced the Mac app on September 17. TechCrunch's September 23 coverage describes the newer demonstrations. Having an app does not establish access to every demonstrated permission or feature.

Glasses, avatars and Charm are relevant to the product's direction. Judge present-day value by the tasks your own account can perform.

Muse invite codes: eligibility versus referral rewards

There is a verified referral program, but its announcement is not a universal access policy. In its September 16 post and reply, Muse says the inviter and friend each receive 1 billion Muse tokens when the friend joins and redeems the code. The official reply says to redeem it in Settings within 48 hours of joining.

That sequence—join, then redeem—describes a reward. It neither proves that everyone needs a code to enter nor guarantees unrestricted access in every country.

The official starting point is muse.ai/join, which presents iOS, Android and web options. The web home page offers sign-in or registration. We did not verify any account's eligibility. The launch announcement described a US rollout; that historical statement is not a complete country list for September 25. A localized page also does not prove local availability. The terms require users to be at least 18, or the higher minimum age applicable locally.

The referral post does not establish how many purchases or tasks a billion tokens will cover, or that the credit can be transferred to the Muse Spark API. Judge the allowance through actual usage rather than an unsupported conversion.

Meta describes free use and subscription plans, and the US App Store listing shows a free download with in-app purchases. We could not recheck exact current plan prices in a readable first-party price list. Use the limits, prices and terms shown to your own account.

Reported task results: useful coordination, consequential errors

TechRadar's Eric Hal Schwartz described shopping research and handling an electricity-company email. Yahoo Finance's Daniel Howley, writing on September 23, reported moving quotes and a restaurant booking, while documenting failures too.

These are small sets of external observations, not a controlled benchmark with repeated trials. They do not support a numerical success rate.

The strongest case for Muse is a task where the goal is clear but coordinating each step is tedious. For something you can finish in a few clicks, briefing the agent, waiting and reviewing may take longer. An incorrect detail also becomes more costly when another action depends on it. Check the reservation, quote, document or activity record—not just a message saying the job is complete.

Privacy and permissions: the boundaries that matter

Useful delegation requires context. The relevant questions are what Muse can access, how that information is used and what happens when you withdraw access.

Meta's security explanation describes an isolated cloud environment, separation of credentials from the task-running agent, and a separate Sentinel layer controlling external actions. When approval is required, it uses explicit interface controls. These are vendor descriptions; we did not independently audit the security architecture.

• Training has a separate control. The privacy policy effective September 17 says interactions may improve models. The setting is on when you first use Muse and can be disabled under Data controls. The policy says changes also apply to past interactions. Interactions can include connector data used by tools; this is not a promise to reverse training that has already occurred.
• Disconnecting does not erase prior context. New exchanges stop, but earlier information may remain in memory or conversations. Deleting a chat may leave learned information behind. Review memory and the available forget or reset controls as appropriate.
• Credential isolation is not complete provider exclusion. Meta distinguishes current Secure VM from planned Confidential VM. Current arrangements still allow access needed to operate, support or secure the service. The planned mode aims to prevent provider access cryptographically; we did not verify a general release.
• The advertising statement has limits. Meta says conversations and VM data are not shared with its advertising systems. It also says browsing or other actions performed on your behalf can indirectly influence advertising.

Start by granting the information and permissions needed for one defined task. Expand the role only after seeing how Muse interprets your instructions and presents results. This recommendation follows from the documented controls and reported mistakes; it is not a blanket security verdict.

Three useful trials after you get access

These are suggested trials, not tests we ran.

Compare public information with clear constraints. Ask for three service providers, their coverage, price sources and booking conditions, with links and the date checked. Inspect outdated information and uncertainty before authorizing contact.

Advance one bounded task to a draft. Use a specific email and request the points that need confirmation plus a proposed reply. Check dates, recipients and commitments, and whether the agent stops at the agreed approval point. Deduct correction time from the time saved drafting.

Try a continuing task with an end condition. Ask it to monitor a public product price for a defined period and notify you only below a threshold. Check whether it remembers the threshold, avoids repetitive alerts and stops when cancelled.

Together, these probe information reliability, execution and continuity. Count briefing, waiting, review and rework when comparing the experience with your current method.

Who should try Muse now?

If scattered information, repeated coordination and follow-ups consume your day, and you are willing to confirm significant actions, Muse has enough concrete examples to justify a trial.

The benefit is less clear for quick questions or work depending on unavailable connectors and forthcoming devices. If material cannot be processed in the cloud, assess the current service; a future Confidential VM plan does not satisfy a present requirement.

Our judgment is that Muse should be evaluated by how much real work it can keep moving. The Connect updates make the direction clearer. Your available features, completed results and supervision time determine whether it earns a place in your routine. Start with an existing chore rather than inventing work just to spend referral tokens.