Meta Muse Review 2026: Features, Pricing, Limits, Security, API and Alternatives

Meta Muse is a personal AI agent that Meta launched on 8 September 2026. Rather than answering questions, it does things: it reads and sends email, books restaurants and travel, fills in forms, compares prices and checks out with a one-time card, all inside a private virtual machine that Meta calls the Muse Secure VM. It is free for most people, with Power ($20 a month) and Maximum ($100 a month) plans for heavier use, and it runs on Meta's own Muse Spark 1.3 model, which developers can also use through the Meta Model API. This guide covers what Muse can do today, what it costs, how the security model works, where it falls short, and how it compares with rivals from OpenAI, Anthropic, Google and others.

What is Meta Muse?

Muse is Meta's answer to the "personal agent" category that OpenAI, Google, Anthropic and xAI have all entered in 2026. Meta describes it as "the world's first personal AI agent built for everyone", and the pitch is that once you share a goal, Muse "helps you develop a personalised plan and coordinate your time and resources, then advances the work on its own". In practice it is a chat-style app, closer to a messaging thread than a chatbot, where you delegate tasks and Muse comes back when it needs a decision or has finished.

Three things distinguish it from a normal assistant. First, it keeps working after you close the app, so a request to "find me a cheaper phone plan" can take hours and involve dozens of web pages. Second, every user gets their own dedicated virtual machine in Meta's cloud, with its own browser and a separate gatekeeping agent called Sentinel that must approve anything Muse tries to send to the internet. Third, it is deeply connected to Meta's own apps: it can turn a recipe you saved on Instagram into a grocery list, or use what you have posted on Facebook to understand your preferences.

Muse is built on Muse Spark 1.3, the flagship model that Meta Superintelligence Labs released on 2 September 2026 and that also powers Meta AI in WhatsApp, Instagram and Facebook, the Muse Code coding agent and the Meta Model API. The consumer product is the subject of this review; developers will find the model and API covered in the API sections. Meta's chief AI officer Alexandr Wang has framed Muse as the first step towards the "personal superintelligence" Mark Zuckerberg described in his August 2026 essay.

Why should someone use Meta Muse?

The strongest reason is price. Meta expects "the vast majority of users" to stay on the free tier, which is unusual: Google's comparable Gemini Spark requires a $99.99 AI Ultra subscription, OpenAI's ChatGPT Work needs at least a $20 Plus plan, Anthropic's Cowork needs Pro, and xAI's Grok Bot starts at $100. If you want to try delegating everyday errands to an agent without paying, Muse is currently the most accessible option among the big platforms.

The second reason is the security architecture. Meta has clearly designed Muse for the trust problem that follows the company around. The agent never sees your passwords or card numbers; credentials sit in a secure store inside the VM and are used without being exposed to the model, and purchases go through Stripe Link with single-use virtual cards. Every outbound action passes through Sentinel, and you can set per-app permissions, for example read-only access to your inbox but no permission to send. Whether that is enough to trust Meta with your email is a personal decision, but the design is more considered than most agents on the market.

The third reason is memory that reaches into places other agents cannot. Muse can draw on your Instagram saves, Facebook activity and WhatsApp conversations (only where you grant access), which makes its unprompted suggestions more relevant than an agent that only sees your calendar. If you live inside Meta's apps, that is a real advantage; if you do not, it is largely irrelevant.

You should probably not use Muse if you are outside the United States (it is not yet available), if you need team or business workflows (it is designed for personal accounts), or if you want a professional-grade agent for coding or document production, where ChatGPT Work, Claude Cowork and Muse Code itself are better fits.

Key features

Task execution across your accounts

Muse connects to Gmail, Google Calendar and Google Workspace, Ticketmaster, OpenTable, Spotify, Apple Health, Peloton, Plaid (for bank and card data), Function Health, Facebook and Instagram, with more connectors promised. It can draft and send email, resolve calendar conflicts, book tables and tickets, track spending and audit subscriptions, and build workout and meal plans from your health data.

Browser automation for everything else

Where a service has no API, Muse opens a visible browser inside its VM and works through the website like a person: filling forms, comparing prices, negotiating with a customer-service chat, or completing a checkout. You can watch the browser window while it works and take over if needed. 1Password integration lets it sign in to sites you already use.

Goal planning and background work

Give Muse a larger goal, such as setting up a small business or training for a race, and it breaks it into a plan, schedules the steps and continues in the background, returning when something changes or it needs approval. Meta says it can juggle several workflows at once, a capability Muse Spark 1.3 was specifically trained for.

Approval gates and audit trail

Before spending money or sending a message, Muse pauses for your approval and shows what it intends to do. Every action is logged and visible to you afterwards. You can loosen or tighten the approval rules per app.

Payments with purchase protection

Checkout uses Stripe's Link with a one-time-use card number generated for each purchase, so merchants never see your real card. Purchases are covered by Link's protections for damaged goods, price drops and returns. Shop Pay support is coming soon.

Memory and personalisation

Muse remembers preferences, dietary restrictions, recurring commitments and the tone you like, and makes unprompted suggestions based on them. You can tell it to forget specific things, and you can choose a name, avatar and communication style for your agent.

Multi-platform access

Muse is available on the web at muse.ai, in dedicated iOS and Android apps, and inside WhatsApp, where you can message it like a contact. Support for Meta's AI glasses is promised but has no date.

How does Meta Muse work?

Each Muse user gets a dedicated virtual machine in Meta's cloud. Inside it run three things: the Muse agent itself (Muse Spark 1.3 plus conventional agent software), a secure store holding your connected-account credentials, and a browser. Alongside, at the system level, sits Sentinel, a separate agent with one job: nothing Muse does reaches the internet unless Sentinel approves it. Sentinel enforces the per-app permissions you set and asks you before anything sensitive happens.

When you give Muse a task, the model plans the steps, decides whether each one can be done through a connector API or needs the browser, and starts executing. Because the VM persists, work can continue after you close the app, and Muse pushes a notification when it needs a decision. Approvals, purchases and sent messages are recorded in an audit trail you can review.

On the privacy side, Meta states that Muse conversations and VM data are not shared with its advertising systems, that Muse cannot see the passwords or payment methods it uses, and that you can opt out of your interactions being used to train Meta's models. Later in 2026 Meta plans a Muse Confidential VM in which the entire machine, conversations included, is encrypted with keys only you hold, so that Meta itself cannot read it. That version was not available at launch.

How to use Meta Muse step by step

1. Install or open Muse. Download "Muse from Meta" from the App Store or Google Play, go to muse.ai on a computer, or message Muse in WhatsApp. You need to be in the US and 18 or older.
2. Sign in. Muse uses your Meta account (Facebook, Instagram or Meta Accounts Center). During launch week some users needed an invite code from the earlier closed alpha; most new sign-ups now proceed directly.
3. Add a payment card to unlock the free tier. Meta asks for a card even on the free plan; you are not charged unless you upgrade or authorise a purchase.
4. Personalise your agent. Choose a name, avatar and how chatty or formal you want it to be.
5. Connect apps. In Settings, link Gmail, Calendar, Instagram, Plaid, Spotify and so on. For each connection choose the access level, for example read-only versus send, and you can disconnect at any time.
6. Give it a first task. Start small: "Find the three cheapest direct flights to Denver next Friday and hold them for me" or "Go through my subscriptions and list anything I have not used in three months".
7. Approve or adjust. Muse shows its plan and stops before sending or paying. Approve, edit, or tell it to take a different approach. Watch the usage meter, which shows how much of your weekly allowance remains.
8. Review the audit trail. After a task completes, open the activity log to see every action, and use "forget this" for anything you do not want retained.

Who can use Meta Muse?

At launch Muse is available only in the United States and only to adults aged 18 and over. Meta has said expansion to other countries is planned but has given no dates; European users in particular should expect a delay given regulatory review of agents that act on personal data. The Android listing shows a 12+ content rating, but Meta's own terms require 18+, and the iOS app is rated 18+.

Muse is designed for individuals and personal accounts. There is no team, family or business tier, no admin console and no shared agent. Small-business owners can use it for their own errands, but it is not a replacement for a work automation platform. You do need a Meta account, so people who have deleted Facebook and Instagram will have to create one.

Practical use cases

Inbox and calendar triage. With read access to Gmail and Calendar, Muse summarises what needs attention, drafts replies for approval, and resolves double bookings by proposing alternatives to the other party.

Dining, tickets and travel. The OpenTable and Ticketmaster connectors let it book a table or seats directly, and the browser handles airlines and hotels. Ask it to watch for a price drop and it will keep checking in the background. Note that some venues have pushed back: the reservation platform Resy has asked users not to point agents at its site.

Subscription audits and bill reduction. Via Plaid, Muse can see recurring charges, flag unused services and, with approval, cancel them or negotiate a lower rate through a provider's chat.

Shopping and deal hunting. It compares prices across retailers, sets deal alerts, and checks out with a single-use Link card once you approve.

Health and fitness plans. With Apple Health, Peloton or Function Health connected, it builds workout and meal plans around your actual data and adjusts them week by week.

Life admin from social content. Saved Instagram recipes become grocery lists; an event you were tagged in on Facebook becomes a calendar entry with travel planned.

Longer projects. Registering a business, planning a wedding or preparing a move can be handed over as a goal; Muse plans the phases and works through them, checking in at each decision point.

Free version and its limitations

Muse's free tier is the product's main selling point. Meta says it is "free for most of what people need", and Alexandr Wang has said most users should be able to do everything they need without paying. The free plan includes the full agent, all connectors, the Secure VM, the browser and payments; it differs from paid plans only in volume.

The catch is that Meta has not published a stable table of allowances. Launch coverage reported a free ceiling of roughly 100 million tokens per week, which is generous for chat but can be consumed quickly by long browser sessions, since every page the agent reads counts. The app shows a usage meter so you can see how much of the week's allowance remains. The free tier also requires a payment card on file, and Meta uses free-tier interactions to improve its models by default, which you can switch off in settings. Finally, the free plan is subject to the same US-only, 18+ restrictions as everything else.

Pricing and subscription plans

Prices are in US dollars, billed monthly through the web or as in-app purchases on iOS and Android. There is no annual discount at launch, no family plan and no business tier. Paid plans add capacity rather than features: connectors, security and payments are identical on all three. For comparison, ChatGPT Plus is $20, Claude Pro is $20, Perplexity Pro is $20, Google AI Ultra with Gemini Spark is $99.99, and xAI's SuperGrok Plus with Grok Bot is $100, so Muse's Power tier is priced with the market and its Maximum tier undercuts Google and xAI slightly while offering far fewer publicly stated guarantees about what you get for it.

Tokens, credits and usage limits

Muse meters usage in tokens rather than tasks or credits, which makes it different from Perplexity Computer (credits) or Manus (credits per task). Tokens are consumed by your conversation, by the model's reasoning, and above all by what the agent reads while working: a long browser session that loads twenty pages can use far more than a hundred chat messages. The weekly allowance resets automatically, and the in-app meter is the only official indicator of where you stand.

As of this review, Meta has published no exact figures for the Power and Maximum tiers, only that they are "higher" and "highest". That is a notable gap compared with competitors that state their allowances, and we would advise trying the free tier for a week and watching the meter before deciding whether $20 or $100 is justified. There is currently no way to buy extra usage without upgrading a tier.

API availability and features (Muse Spark)

The consumer Muse agent has no public API: you cannot script or embed it. What developers can use is the model underneath, Muse Spark 1.3, through the Meta Model API, which Meta opened in public preview on 9 July 2026 alongside Muse Spark 1.1 and updated for each release since. The API is OpenAI-compatible, meaning you can point the OpenAI SDKs at Meta's endpoint with a different base URL and key, and it is also available through third-party gateways such as OpenRouter and on AWS.

On benchmarks, Muse Spark 1.3 is strongest in coding and long-context retrieval: 75.4 on DeepSWE v1.1 (ahead of GPT-5.6 Sol at 73.0 and Claude Opus 5 at 74.0), 88.8 on Terminal-Bench 2.1 (tied with Sol), 59.4 on SWEAtlas, and 98.1 on MRCR retrieval at 512K to 1M tokens against 73.8 for Sol. On agent benchmarks it is competitive but not leading: 66.9 on OSWorld 2.0 versus 68.3 for Opus 5, and 64.9 on JobBench versus 65.7 for Opus 5. Artificial Analysis rates the xhigh variant at an Intelligence Index of 61 with throughput of about 182 tokens per second. Meta also reports improved resistance to prompt injection and better calibration on irreversible actions, both of which matter for an agent model.

How to obtain an API key

1. Go to developer.meta.com/ai and sign in with a Meta developer account (the same account used for Meta's other developer products; create one if needed).
2. Open the Meta Model API section and accept the developer terms. Read the data-use choice carefully: the Standard tier does not train on your prompts, while the Contributor tier gives Meta permission to use your prompts and completions to improve future models in exchange for much lower prices.
3. Create an app or project, then generate an API key from the keys page. Copy it immediately and store it in a secrets manager or environment variable.
4. New accounts receive a one-time $20 credit to prototype with; add a payment method to continue on pay-as-you-go once it is used.
5. Confirm the model ID for Muse Spark 1.3 in your account's model list, since Meta has noted that model availability and tiers now need account-level verification after the 1.3 rollout.

API pricing

Meta has held the price steady across 1.1, 1.2 and 1.3, which Wang describes as "aggressive". At $1.25 / $4.25 the Standard tier is well below GPT-5.6 Sol ($5 / $30), Claude Opus 5 ($5 / $25) and GPT-6 Astra ($10 / $50), and roughly in line with Gemini 3.1 Pro ($2 / $12) and Grok 4.6 ($2 / $6). The Contributor tier is over 90% cheaper again, and Meta says adoption is in the "meaningful double digits" among developers, but it is only appropriate for data you are entitled to share. Pricing for the "max" reasoning variant had not been published at the time of writing, and Meta's page describes the API as a public preview with no change-notice commitment, so re-check before committing a budget. Prices are third-party-corroborated but Meta's own pricing page renders client-side; verify against your account dashboard.

How to integrate and use the API

Because the Meta Model API follows the OpenAI request format, integration is usually a matter of changing the base URL, key and model name in code you already have. Check the model ID and base URL in your developer dashboard, since Meta's documentation identifies these per account.

# Python — server-side only
from openai import OpenAI
client = OpenAI(
    api_key=os.environ["META_MODEL_API_KEY"],
    base_url="https://YOUR-ENDPOINT-FROM-DASHBOARD/v1",
)

response = client.chat.completions.create(
    model="muse-spark-1.3",          # confirm exact ID in your dashboard
    messages=[
        {"role": "system", "content": "You are a careful assistant that plans before acting."},
        {"role": "user", "content": "Draft a three-step plan to migrate our newsletter to a new provider."}
    ],
    extra_body={"reasoning_effort": "high"},
)
print(response.choices[0].message.content)

Practical notes from early users: the model does not expose its reasoning trace, which makes debugging agent decisions harder than with some rivals; use structured output and explicit tool schemas to keep it on track. Prompt caching at $0.15 per million makes long system prompts and documents cheap to reuse. If you want an agent product rather than raw model access, Muse Code (Meta's coding agent, installable on macOS and Linux) uses the same model and tiers. For a broader look at agent APIs see our OpenAI API guide.

Limitations and disadvantages

The most significant caveat is maturity. Muse is one day old at the time of writing. Testers reported connections dropping mid-task and, more seriously, cases where guardrails were bypassed; Reuters reported that internal testing before launch produced "mixed results" including "serious security concerns involving personal data". Meta has said it has "meaningfully increased" safety investment but has not had to pause the product. Any agent that drives a browser is also vulnerable to website changes and errors, and the more accounts you connect, the larger the consequences of a mistake.

Transparency is the second problem. Meta has not published the free tier's cap, the paid tiers' caps, or a detailed data-handling document for the VM. The promise that conversations stay out of the ads system is a policy statement, not an audited control, and the Confidential VM that would make it enforceable technically is still months away.

Finally, Muse is narrower than it first appears. It is excellent at consumer errands but not designed for professional output: it will not build you a spreadsheet model or refactor a codebase the way ChatGPT Work, Claude Cowork or Perplexity Computer will. Judge it as a digital personal assistant, not a work agent.

Ethical, privacy and security considerations

Data and advertising. Meta states that Muse conversations and VM contents are not used by its advertising systems and that there are no ads inside Muse. Interactions may be used to improve Meta's AI models by default; you can opt out in settings and can instruct Muse to forget specific information. Given Meta's regulatory history, including the 2019 FTC settlement over privacy practices, many users will want to wait for the Confidential VM before connecting financial or health accounts.

Credentials and payments. Muse does not have visibility into passwords or payment methods; they are held in the VM's secure store and used by the browser without being exposed to the model. Purchases use single-use Link cards. Still, granting an agent your bank data via Plaid or your inbox via Gmail is a meaningful expansion of access; start with read-only permissions and widen them only as trust builds.

Prompt injection. An agent that reads web pages and emails can be manipulated by text hidden in them. Muse Spark 1.3 was trained for stronger resistance to prompt injection and Sentinel adds an independent check, but the launch-week guardrail reports show the risk is real. Keep approval gates on for sending and spending.

Third-party terms. Some websites prohibit automated access, and at least one reservation platform has asked users not to use agents on it. Muse's browser use may breach those sites' terms even if it is legal, and accounts could be restricted as a result.

Minors. Muse is 18+ only. Meta AI's other surfaces have separate teen protections; Muse is not one of them.

Watermarks and commercial-use restrictions

Text that Muse writes, such as emails, plans and summaries, carries no watermark. Images created by Meta's Muse Image model in Meta AI carry visible and invisible AI markers; Muse the agent does not generate images itself at launch. Under Meta's AI terms you own the content you create with Muse and can use it for personal or commercial purposes, subject to the terms and to third-party rights. The main practical restriction is that Muse is a personal product: the terms are written for individual use, and there is no licence covering a team or a business acting on behalf of clients. Developers using the Muse Spark API own their outputs under the Meta Model API terms, with the Contributor tier's data-sharing caveat described above.

Sign-up, sign-in and login instructions

You sign in to Muse with a Meta account. On the web, go to muse.ai, choose Continue with Facebook or Instagram (or your Meta Accounts Center login), confirm you are in the US and over 18, and add a payment card. In the mobile apps the flow is the same after installation, and in WhatsApp you start a chat with the Muse contact and link your Meta account when prompted. Two-factor authentication on your Meta account protects Muse as well, so enable it under Accounts Center → Password and security before connecting anything sensitive.

If you have deleted Facebook and Instagram, you will need to create a new Meta account. Because Muse can act inside your other accounts, treat the Meta login as a master key: use a unique password, a hardware key or authenticator app, and review the connected apps list regularly. To stop Muse entirely, disconnect all apps in Settings and delete the agent; Meta says VM data is removed with it.

iOS and Android app availability

Official apps for both platforms launched alongside the service on 8 September 2026. They offer the full experience, including the visible browser window, approvals, notifications when a background task needs you, and the usage meter, and they are the only official mobile channels apart from WhatsApp. Do not install Muse from third-party APK sites.

Muse from Meta for iOS

Published by Meta Platforms, Inc., the iPhone app requires iOS 18.0 or later, is about 114 MB, is rated 18+ and supports 39 languages, although the service itself is US-only for now. It offers in-app purchases for the Power and Maximum plans. Download Muse from Meta on the App Store.

Muse from Meta for Android

The Android app (package com.facebook.aura) was published on 8 September 2026 and had passed 1,000 downloads within a day with early five-star reviews too few to be meaningful. Device requirements are not yet listed; expect a recent Android version with Google Play services. Get Muse from Meta on Google Play.

Muse is separate from the Meta AI app, which remains the chat-and-image assistant and is reviewed in our Meta AI guide.

Meta Muse and Muse Spark videos

A news explainer on Meta's Muse Spark launch and two independent hands-on tests of the Muse Spark 1.3 model that powers the agent.

Top 10 Meta Muse alternatives

These are agentic assistants that act on your behalf, not plain chatbots. Prices and limits were checked against official pages on 9 September 2026 and change often; confirm on each vendor's site before subscribing. Text output from all of them is unwatermarked; watermark notes refer to generated images.

Comparison table

Final verdict: is Meta Muse worth using?

Main strengths. Muse is the most accessible personal agent from a major platform: a free tier with every feature, a security design that keeps credentials away from the model and puts a gatekeeper between the agent and the internet, background execution with approval gates, and connectors that reach into Meta's own apps as well as Gmail, Calendar, Plaid and health services. The Muse Spark 1.3 model underneath is capable and, for developers, cheap.

Main weaknesses. It is one day old, US-only and 18+, with unpublished usage caps, a Meta account and card required even for free use, reports of early reliability and guardrail problems, and a privacy promise that rests on policy rather than the user-encrypted VM still to come. It is a personal errand-runner, not a work agent.

Who should use it. US adults who already live in Meta's apps and want to try delegating everyday tasks such as email triage, bookings, subscription audits and shopping without paying. Anyone curious about personal agents will learn a lot from the free tier.

Who should consider an alternative. Users outside the US; anyone unwilling to give Meta access to email or bank data; people who need professional deliverables (ChatGPT Work, Claude Cowork, Perplexity Computer); technical users who want full data control (OpenClaw); and households that want voice-first errands (Alexa+).

Does the free version provide enough value? Yes, and it is the best reason to try Muse. Meta itself expects most people never to pay.

Does the paid version offer good value? Not yet provable. Power at $20 is priced with the market, but without published caps you cannot compare it with Perplexity's credit table or Manus's daily allowance. Maximum at $100 only makes sense for someone running long browser tasks continuously, and even then a week on the meter is the only way to know.

Is Meta Muse worth it? As a free experiment, yes: it is a well-designed, capable agent that costs nothing to try, and the security architecture is ahead of most rivals on paper. As a paid subscription, wait: let the launch-week bugs settle, watch for the Confidential VM, and see whether Meta publishes the limits you would be paying for. Start with read-only permissions and expand from there.

Sources: Meta Newsroom announcement "Introducing Muse" (8 September 2026), Meta AI Research post "Introducing Muse Spark 1.3" (2 September 2026), Meta for Developers (Meta Model API), the Apple App Store and Google Play listings for Muse from Meta, reporting by Axios, TechCrunch, CNBC and 9to5Mac, and each alternative vendor's official pricing and documentation pages, all accessed 9 September 2026.