Muse Review: Can Meta’s AI Get Things Done? — November 2026

Muse is worth watching, and eligible users have a sensible reason to try it. Paying for it depends on whether it repeatedly finishes work that matters to you. Its strongest proposition is the combination of cross-service execution and follow-up. Our reservations concern availability, the effort of managing delegated access, and the lack of independently verified completion data for the consumer product.

Consider an event invitation that needs to become a calendar entry, a travel decision, a document checklist, and a reminder to somebody else. A helpful answer can explain every step while leaving the entire job in your hands. An agent promises to take responsibility for more of that sequence. That changes the review question: a charming conversation matters less than whether the final calendar entry, file, and notification match your intent. This example illustrates our criteria; it is not a task we ran.

1. What Muse is, and who makes it

Muse is a personal agent from US company Meta, publicly introduced on September 8, 2026. Its consumer interfaces include mobile, web, and WhatsApp, with work taking place in a cloud environment. Treat it as an account-based hosted service involving permissions and external providers. Downloading a related model would not reproduce that complete service.Meta: introducing Muse · Muse product page

The current Muse for Small Business page specifies adults in the United States and Canada. Meta’s subscription help also says availability and benefits vary by location and account. Readers elsewhere should separate interest in the technology from access to the actual product. We have not established equivalent reliability for Chinese conversations, Chinese websites, or local commercial workflows.Muse for Small Business and availability · Meta: Muse subscription plans

Meta: introducing Muse · Muse for Small Business and availability · Meta: introducing Muse Wonder · Meta: additional Muse tokens

Our assessment is that Muse addresses fragmented work. People often know what they should do; the problem is finding the time to collect inputs, open several services, make changes, and remember to follow up. A common point of delegation becomes attractive when those steps recur. When a task happens rarely and takes two minutes to finish manually, explaining it to an agent may consume the potential saving before the agent starts.

2. Separate the app, the models, and the coding product

Muse is the personal-agent product. Spark and Wonder are model families, while Muse Code is the developer-facing coding tool. Meta’s catalog currently identifies Spark 1.3 as the latest Spark version and lists muse-wonder-1.0 for Wonder. It explicitly distinguishes Wonder Full from the normal Wonder model: selecting maximum reasoning effort does not select Full.Meta model catalog

Muse product page · Meta model catalog · Meta: introducing Muse Wonder · Meta: privacy, safety and security

We found no single public stable-release number for the whole Muse app, or a routing table covering every account and feature. We leave those gaps open. A model catalog establishes what developers can access; an announcement establishes what a company released or intends to release. Neither tells an individual exactly what is running behind a particular consumer task. The relevant purchase information is the benefit description presented to that account.

This distinction also prevents misleading attribution. A task can fail because an instruction was misunderstood, a connector lacked access, a website changed, a session expired, or approval was pending. A stronger model cannot automatically remove all those constraints. Equally, a successful appointment booking says little about the model’s performance on unrelated reasoning tasks. Version information helps most when it explains which practical jobs could improve.

3. Core capabilities: completing work over time

A browser extends reach and raises the acceptance bar

Meta’s browser help lists appointment booking, purchases, forms, and price monitoring. Users can inspect the browsing session, take control, or stop it. We see that as broader reach coupled with a need for explicit acceptance criteria. Counting successful clicks would tell us less than checking whether the requested outcome actually exists.Meta: browsing and taking control

For a hypothetical restaurant booking, finding a suitable venue is only the beginning. The date, party size, contact details, and cancellation conditions all affect whether the result is useful. One wrong detail can send the user back into the process. A confirmation that can be checked against the request is therefore a better deliverable than a reassuring statement that everything is handled. This is the ordinary discipline of delegating work, not a demand that consumers learn technical prompting.

Memory needs correction, and persistence needs an ending

The design essay describes persistent context, a main conversation with side chats, background follow-up, and rich outputs. These features make an ongoing relationship possible. Their appeal is less repetition; their burden is keeping the agent’s understanding current. An assistant that remembers an outdated preference can be more inconvenient than one that asks again.How We Designed Muse

Suppose your usual travel day changes from Friday to Thursday. Helpful memory should incorporate that change rather than confidently keep applying the old schedule. Proactivity then depends on judgment: a material change or blocked decision may deserve a notification, while a reworded status update probably does not. We did not test those situations. They explain why memory capacity alone is an incomplete measure of personal-assistant quality.

Persistence also creates obligations between related tasks. Canceling a plan should raise questions about its reminders, unfinished drafts, and dependent actions. Starting work proves that delegation began; stopping the correct work and preserving the useful result prove that the job ended. Those lifecycle details are particularly important when the user is not watching every step. We cannot infer them from a polished launch demonstration.

Read the evaluations according to their actual metrics

The Wonder model card separates desktop interaction, memory, persistence, and long coding work. OSWorld V2.1 uses mean partial reward as its main metric, with binary completion secondary. MemoryBench and FocusBench are internal evaluations. Some SWE-Marathon results use different agent harnesses. None of these measures establishes a consumer Muse app completion rate.Muse Wonder model card and evaluation methods

Partial progress and a finished transaction can have very different value. Reaching a checkout page, entering the correct address, and receiving the right order confirmation are separate states. Evaluation numbers can look comparable while answering different questions. We would rather leave a missing product-level measurement visible than combine memory, coding, and browsing into a percentage chart that suggests a common standard where none was established.

For this product, stronger evidence would describe the requested outcome, permissions, external service conditions, human interventions, and final state. We did not obtain an independently reproduced long-running Muse app evaluation covering those factors. Public model research supports taking the design seriously; it does not remove the need to verify a specific consumer job. Our recommendation remains conditional partly for that reason.

4. Tools and integrations: access is the beginning

Connectors link Muse to services, while skills provide particular capabilities. The artifacts help page names formats including PDF, DOCX, XLSX, CSV, and interactive web outputs. Together these can turn information from connected services into something usable. Format support alone does not establish correct formulas, complex document layout, or complete application behavior.Meta: Muse connectors · Meta: Muse skills · Meta: Muse artifacts

The small-business page lists integrations such as Notion, Shopify, Slack, Stripe, and QuickBooks. That combination could matter more to an owner than another writing feature because less information would need to be moved by hand. A connector list still does not establish the scope available in a particular account, organizational controls, or a service-level commitment.Muse for Small Business and availability

Disconnecting access has a narrower meaning than removing history. Meta says it stops future exchanges, while previously used information may remain in conversations and memory. Some Meta social accounts in the same Accounts Center can also connect automatically. A user should therefore understand both current access and retained context. Turning a connector off is not a complete erasure mechanism.Meta: Muse connectors

Custom connectors extend reach, but Meta says it does not review them or their information use. We would favor established integrations for recurring everyday work and treat custom connections as an additional maintenance decision. A service can change its interface, permissions, or behavior after setup. AI assistance may reduce implementation effort without eliminating responsibility for what was connected and how that connection behaves.Meta: Muse connectors

5. Permissions and safety: understandable delegation

Meta describes separating the execution environment from credential and permission services, with Sentinel controlling connector actions and network egress. We see value in moving authorization outside the core model. We did not audit the implementation, and a documented architecture is not proof that every product feature is secure.Meta: How We Built Safety Into Muse

The approval help describes one-time, task, site, and continuing grants, plus activity records. Purchase guidance says checkout always requires approval, with availability depending on location and merchant. An authorized action can still be wrong. A person approving a purchase must assess the actual item, amount, and conditions, not merely the fact that the product displayed a confirmation.Meta: guidance and approval · Meta: Muse payments

We would judge a permission interface by how quickly a user can understand who receives what, what will change, and whether money is involved. An ambiguous request to continue adds friction without necessarily supporting a decision. Excessive low-value prompts can also encourage habitual approval. The useful balance is specific consent at meaningful boundaries and clearly bounded access for routine work.

Undo is another boundary. A message already sent, information already disclosed, or data already processed by another provider may have effects outside the assistant’s control. Meta’s own guidance notes that sending email cannot be reversed in the way a browsing task can be stopped. A final human check on consequential communication is a normal part of delegation, even when the preparation was automated.Meta: guidance and approval

6. Privacy: advertising, training, and sharing differ

Meta says conversations and VM data are not shared with its advertising systems. Its data help separately says the model-improvement setting starts enabled and can be switched off under Data controls. These are different policies. A reader who wants to experiment without contributing interactions to training should check that setting rather than infer a training exclusion from the advertising promise.Meta: privacy, safety and security · Meta: data controls and memory

The architecture post also identifies indirect advertising effects from agent browsing. It says the current design does not cryptographically prevent operational access by Meta, while Confidential VM is future work. We did not verify universal availability of that confidential environment. We consequently do not treat provider-inaccessible storage as today’s default.Meta: How We Built Safety Into Muse

Meta: privacy, safety and security · Meta: data controls and memory · Meta: Muse artifacts · Meta: How We Built Safety Into Muse

A recent sharing issue matters here. Claroc reported finding over 6,800 shared hostnames through passive DNS and opening a sample of several hundred pages. That is not a count of compromised accounts. An October 23 TNW follow-up quoted a Muse team member saying the artifact-link issue was fixed, sharing restored, and old artifacts assigned new URLs. We checked the research and follow-up report, not the fix itself. We do not present the earlier pause as a confirmed continuing state.Claroc: original GhostShares research · TNW: October 23 follow-up on the sharing fix

The lesson concerns the sharing boundary, not a conclusion that every private conversation became public. Credential isolation can be thoughtfully designed while another product surface still exposes information in an unexpected way. Before sharing an output, its title, personal details, and business material deserve their own review. At that point the question changes from what the assistant knows to what an external recipient can access.

Meta documents memory editing, export, and resetting, and warns that deleting a message can leave information the agent learned. We consider exit costs part of everyday product quality. Canceling a bill, stopping active tasks, revoking access, and deciding what happens to retained information are different concerns. The more useful a long-lived assistant becomes, the more important it is that a person can understand those choices.Meta: data controls and memory

7. Pricing: distinguish the plan from the completed job

The official help page lists capped free access, Power at $20 per month with 500 million weekly Muse tokens, and Maximum at $100 with 3 billion. Subscriptions renew monthly; benefits vary by account and region. The page does not publish a numerical free allowance, so we do not substitute referral offers or early promotional numbers.Meta: Muse subscription plans

Meta: Muse subscription plans

The last column multiplies the monthly price by twelve. It is not an annual-billing offer, and excludes tax, extra usage, and shopping expenditure. Maximum costs five times as much for six times the stated weekly allowance. At full utilization, that is roughly 16.7% less per unit of listed allowance. It is not a promise of cheaper individual tasks, sixfold speed, or sixfold productivity.

Meta now documents one-time additional-token purchases for free and subscribed users. Purchased extras do not expire; included weekly subscription usage is subject to non-rollover rules. Consumption varies by request, model, tools, and context, with no guaranteed output quantity or quality per allowance. We did not obtain a uniform public top-up price and do not invent one.Meta: additional Muse tokens

A hypothetical time budget makes the buying decision clearer. At a personal time value of $20 an hour, Power needs to save one net hour a month to offset its fee; Maximum needs five. Net means after briefing, approvals, checking, and rework. If an agent reduces a thirty-minute manual job to twenty-five minutes of your own involvement, the saving is five minutes. These are illustrative assumptions, not measured Muse results.

Model API access is a separate bill

Meta Model API pricing

These rates do not buy the whole personal-agent service. Contributor pricing permits training use in exchange for a discount. For an illustrative Standard-mode workload of one million uncached input tokens and 200,000 output tokens, Spark text inference costs $2.10 and Wonder $8.00, before search, tools, or hosting. Those figures cannot establish how many emails a personal subscription will handle.Meta Model API pricing

Our purchase advice starts with recurring work. If capped free access already supports your occasional requests, a larger quota alone is not a reason to upgrade. Power makes more sense after identifying a repeated task with a reviewable benefit. Maximum needs a stronger usage case. Deciding whether a task deserves delegation is usually more useful than deciding which large allowance looks best in a table.

8. Usability: less transfer, more verification

Without operating the product, we cannot rate interaction feel, waiting time, or Chinese conversational quality. We can analyze the proposed workflow. Placing requests, progress, approvals, and outputs close together could reduce switching costs. Initial setup still involves connecting accounts, explaining preferences, defining boundaries, and learning how to judge results. A launch demonstration cannot establish how much of that effort will disappear in an individual’s routine.

The most promising tasks are frequent, bounded, and easy to check. Combining scattered information into a reviewable list or watching a specified condition can fit that pattern. A two-minute personal task requiring ten minutes of explanation does not. This is why broad capability and personal usefulness are different measures: a product can do something without that being the best way for you to get it done.

Migration involves more than importing a file. A person may need to rebuild preferences, recreate active tasks, and inspect what belongs in shared spaces. Persistent context can make an assistant more useful and departure more expensive. We value readable memory, downloadable outputs, and stoppable tasks as selection criteria; we have not established lossless migration for complex workflows.

Small businesses face an additional coordination cost. An owner’s willingness to connect a tool does not settle how every customer record should be used, or which employee can authorize publication. Somebody still supplies the material, somebody checks the output, and somebody approves the external action. Clear roles make automation useful. Leaving those decisions implicit can move the work into a chat transcript without actually reducing it.

9. Muse, Claude, or Manus? Compare where work happens

Computer operation is not unique to Muse. Anthropic describes computer use in Claude Cowork and Code with Dispatch, requiring the desktop app to remain awake and running; the article still labels the capability a research preview. Manus describes a persistent Cloud Computer that can retain files and tools independently of the user’s laptop. These are different working arrangements, not matched performance measurements.Anthropic: Claude computer use and Dispatch · Manus: Cloud Computer

Muse product page · Anthropic: Claude computer use and Dispatch · Manus: Cloud Computer

Start with where your work already happens. Local files and desktop applications may favor collaboration with that computer. Jobs that must continue while it is off favor an independent environment. Fragmented everyday errands make Muse’s organization of goals and services especially relevant. These are editorial judgments about fit, not evidence that one product beats the others on those tasks.

Then compare what each payment actually includes. Model inference, agent usage, a persistent computer, and external services may appear on separate bills. We have not filled the comparison with unverified competitor prices or assumed identical benefits at a similar headline fee. A cheaper entry point that cannot handle the required job may be poor value; a more elaborate environment may create more management work than a casual user wants.

10. Rating and recommendation

Our rating is 7.2 out of 10: conditionally recommended. This is a subjective editorial assessment of current selection value, not a benchmark, satisfaction survey, or measured completion rate. Six dimensions have equal weight. Zero means little practical fit or insufficient information, five means substantial conditions, and ten means strong capability supported by complete public evidence. New testing or product changes could alter the judgments.

The calculation is (8.5 + 8.0 + 8.0 + 6.0 + 7.0 + 5.5) ÷ 6 = 7.1667, rounded to 7.2. We deliberately do not assign a supposed measured reliability score. Task value assesses the practical direction of the documented capabilities. A numerical recommendation should not imply more certainty than the underlying evidence allows.

Eligible users with recurring email, calendar, or cross-site tasks are the clearest candidates, provided they are comfortable reviewing results and managing access. Small-business owners can assess bounded preparation and drafting work. We would not recommend assuming mature Chinese local workflows, unattended consequential execution, default provider-inaccessible data, or universal website coverage simply because a subscription was purchased.

Muse makes the personal-agent proposition concrete and exposes its hardest questions: whether memory stays current, consent stays understandable, work reaches the correct ending, and a person can leave with control of their information. That makes it worth reviewing. For an actual user, the strongest recommendation should remain a recurring job handled correctly with less personal effort. The promise of personal superintelligence cannot substitute for that result.

Sources

Sources checked 2026-11-03. Official sources are prioritized. Sharing remediation is attributed to the linked follow-up report and was not independently retested.

1. Meta: introducing Muse
2. Muse product page
3. How We Designed Muse
4. Meta model catalog
5. Meta: introducing Muse Wonder
6. Muse Wonder research overview
7. Muse Wonder model card and evaluation methods
8. Meta: Muse connectors
9. Meta: browsing and taking control
10. Meta: Muse skills
11. Meta: Muse artifacts
12. Meta: guidance and approval
13. Meta: How We Built Safety Into Muse
14. Meta: privacy, safety and security
15. Meta: data controls and memory
16. Meta: Muse payments
17. Meta: Muse subscription plans
18. Meta: additional Muse tokens
19. Meta Model API pricing
20. Muse for Small Business and availability
21. Claroc: original GhostShares research
22. TNW: October 23 follow-up on the sharing fix
23. Anthropic: Claude computer use and Dispatch
24. Manus: Cloud Computer

This review was generated with the assistance of AI technology.