Meta Muse and Your Website: What the AI Agent Does and How to Handle It

Meta launched Muse on 8 September 2026. It is not a chatbot that answers and stops. It is a personal AI agent with its own computer in the cloud, and it opens websites to book, buy, compare and sort out problems for the person who asked. If you run a shop, a booking site or any business that takes enquiries online, Muse is already a new kind of visitor on your pages.

This guide explains, from Meta's own posts and from Cloudflare's verified-bot registry, what Muse is, how it reaches your site, how to find it in your logs, and how to decide whether to welcome it or keep it out. To check your own logs right now, paste them into the free AI agent log checker. Nothing is uploaded.

What Muse is, in plain words

The Meta launch post describes Muse as "a secure, private personal AI agent that proactively helps with people's goals". People talk to it like they message a friend, in the Muse app or in WhatsApp. They tell it what needs to get done and it takes action, powered by Muse Spark, which Meta calls its most capable model to date.

The important part for a website owner is the way Muse does that work. Meta says it "can open a browser, fill out forms, and negotiate on their behalf". For longer jobs it "keeps working after people close the app, and comes back when something changes or when it needs approval, like before it sends an email or makes a purchase."

So Muse is closer to a human assistant with a laptop than to a search engine. It does not read your whole site. It opens the pages it needs for one task, for one person, and then it leaves.

The key facts at a glance

How Muse reaches your website

Every Muse user gets their own computer in the cloud. Meta's Meta business post on personal agents puts it simply: Muse works "by giving every person their own computer and browser in the cloud, with access to the internet where Muse can run code and use the web under the user's direction to achieve their goals."

That has three effects you can see from your side of the connection.

1. Requests come from cloud addresses, not homes. Muse runs on Meta's infrastructure, so a visit arrives from a data-centre IP rather than a consumer broadband line. Firewalls that distrust data-centre traffic will treat it with suspicion.
2. It uses a real browser. Pages load with their scripts, cookies and forms, like any other visit. Analytics tags normally fire, so Muse sessions can appear in your reports as ordinary visits.
3. There is a gatekeeper on the machine. Meta says a separate Sentinel agent runs on the same VM, and "nothing Muse does reaches the internet unless the Sentinel approves it". That is a control for the user's safety, not something you can configure, but it explains why Muse sometimes stops and asks its user before continuing.

Muse can also use a connector instead of the browser. Meta says Muse "buys things, makes appointments, and resolves support questions through a connector or directly on the business's website". A connector talks to a business's systems directly, so those tasks may never touch your public pages at all.

What the muse- token means

The Cloudflare Radar entry for Muse lists Muse with the operator Meta, the category AI assistant and the user-agent pattern muse-, described as an "automated AI agent client associated with Meta's Muse product". Cloudflare only lists bots that its verification team has reviewed, which is why many Cloudflare-protected sites can already allow or block Muse by its bot identity.

The pattern is a prefix, so you should expect the full string in your logs to be a little longer, for example muse-agent/1.0 or a browser string that contains it. The Meta web crawlers page does not name Muse at all as of 11 October 2026. Until it does, our Muse profile carries the Reported badge, and we suggest you treat a bare muse- string as a claim, not a proof.

Muse is not Meta's crawler

Meta runs several automated clients, and they have different jobs. Mixing them up leads to rules that block the wrong thing. Here is how they compare, using the Meta web crawlers page and Cloudflare Radar.

The practical rule: if you want to stop Meta training on your content, block Meta-ExternalAgent in robots.txt. That has no effect on Muse. If you want to stop Muse, you need a different tool, which we cover below. And if you use Meta business integrations, leave Meta-ExternalTest alone: the Radar entry for Meta-ExternalTest describes it as "used to test external integrations", and blocking it can make an integration fail its checks.

The rules Muse says it follows

Meta has published more about how its agent should behave on other people's websites than most vendors. The Meta business post on personal agents says that when browsing, "Muse behaves like a personal assistant acting for a single person, held to sensible standards." It describes two tests that Muse applies before it acts.

The "one honest person" test

Meta asks: "Would a reasonable, honest person doing this by hand do it this way and at this scale?" Its own examples are useful. A person can refresh a page 500 times an hour, but they would not. A person can make five accounts to redeem a one-per-household coupon, but they should not. A person can write a fake five-star review, but they should not.

The "if every agent did this" test

The second question is: "If every Muse user made this request, would the underlying system still function?" Meta's example is tee times. One person hoarding them is annoying; every agent hoarding them breaks the market. When a request fails a test, Meta says Muse "does what it can, skips what it cannot, and explains the difference to the user."

Approval before consequential actions

Meta says Muse "will seek approval for sensitive actions taken on your behalf like signing in, completing a reservation, or making a purchase." The launch post adds that Muse "has no visibility into people's passwords or payment methods", and that credentials are kept in secure storage so Muse can use them without seeing them.

None of this is a contract with you, and you should not rely on another company's policy to protect your site. But it does tell you what to expect: an agent that pauses at sign-in and checkout, that hands those steps back to its user, and that is designed not to hammer your servers.

What happened with Amazon

On 21 September 2026, Amazon blocked Muse from its retail site. According to the Wikipedia summary of the Amazon dispute, which cites GeekWire's reporting, users who tried to shop through Muse saw a message that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use". Amazon said Meta had not told it that Muse would visit, that the agent did not identify itself while browsing, and that it appeared to store customer credentials. Meta had already said Muse cannot see passwords or payment methods.

Other large retailers went the other way. The same summary notes that Shopify allowed Muse on its storefronts, and Meta's Meta business post on personal agents lists Shopify, Stripe and Walmart among the partners building an open protocol for personal agents with Meta and Sierra.

There are three lessons here for a smaller business.

• Identification is the real argument. Amazon's complaint was that the agent did not say who it was. Since then, Cloudflare has listed Muse as a verified bot with the muse- pattern, which gives site owners a way to recognise it. Whatever you decide, decide it based on verified identity, not a guess.
• Big platforms have reasons you may not share. Amazon runs one of the largest advertising businesses in retail, and agent traffic skips the sponsored results that pay for it. A local furniture shop or a dental clinic has no such conflict. For them, a booking made by an agent is simply a booking.
• Blocking is visible to the customer. When Muse is turned away, its user is told. Meta's own words are blunt: "Turning away a personal agent means turning away the customer behind it."

Should you allow or block Muse?

For most businesses the answer is to allow it on the pages where customers research and act, and to keep your normal protections on the few actions where automation causes harm. Use this table to place your own site.

If you are not sure, start by measuring. Run the AI agent log checker over a week of logs. If Muse shows up on your pricing or booking pages, those are customers trying to reach you.

How to find Muse in your logs

Because Muse uses a real browser, the only reliable mark in a standard access log is its user-agent token. Search for it like this:

grep -i "muse-" /var/log/nginx/access.log | tail -200

grep -i "muse-" /var/log/nginx/access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head -20

The AI agent log checker does the same in your browser and also groups the results by IP, flags whether each operator documents its token, and finds other agents such as Grok Bot (GrokAgent), Hark (HarkBot) and Google-Agent in the same pass. It does not upload anything.

If your site is on Cloudflare, the bot analytics view is even better, because it shows only verified Muse traffic and ignores anyone who copied the string. Filter by the verified bot name rather than by user agent.

How to welcome Muse properly

Allowing Muse is not only about not blocking it. Agents fail quietly on pages that real people can use with a little patience. These are the fixes that matter most, in order.

1. Keep CAPTCHAs off browsing and booking

A CAPTCHA on a search box or a booking form stops the agent until its user steps in. Many users will not. Move human checks to the actions that need them, such as account creation or high-value orders, and use invisible or risk-based modes elsewhere. Our AI Agent Readiness Check shows whether a cloud browser meets a challenge on your key pages.

2. Make prices, stock and hours readable as text

Agents read the page they are given. Prices hidden in images, stock levels that load after a long script, or opening hours inside a map widget are easy for a person and hard for an agent. Put them in plain HTML and mark them up with structured data such as Product, Offer and LocalBusiness.

3. Offer guest checkout and common wallets

Muse pays with Link by Stripe and Meta says Shop Pay support is coming. If your checkout forces account creation, or only accepts one unusual payment method, the agent has to hand the job back to its user. A guest path and a mainstream wallet keep the sale moving.

4. Write forms that explain themselves

Label every field, give clear error messages, and avoid fields that only make sense with a visual cue. An agent filling a form for a person needs the same things a screen-reader user needs, which is why agent readiness and accessibility overlap so much.

5. Do not punish data-centre traffic by default

Muse, ChatGPT's dots and Grok Bot all run on cloud computers. A firewall rule that blocks every data-centre IP blocks all of them, and also many people on corporate VPNs. Prefer rules that look at behaviour and verified identity.

How to block or limit Muse if you must

If you have a real reason to keep agents out, use controls that are enforced, not requested.

Cloudflare

Cloudflare lists Muse as a verified bot, so you can create a WAF custom rule that matches its verified bot identity and set the action to block or managed challenge. Because the rule uses Cloudflare's verification rather than the header, it will not catch people who copy the string, and it will not miss real Muse traffic that sends a longer string.

(lower(http.user_agent) contains "muse-")

Use the user-agent expression only if you cannot use verified bot fields, and remember that it can match unrelated clients.

Other CDNs and servers

if ($http_user_agent ~* "muse-") {
    return 403;
}

Meta publishes no IP ranges for Muse, so there is no address list to block. If your bot manager offers a Muse entry, use it. If not, rate limits on sensitive actions will catch abuse from any agent without naming one.

What about robots.txt?

You can write a group for the token, and it does no harm:

User-agent: muse-
Disallow: /checkout/

But Meta has not said that Muse reads robots.txt, and the standard was written for crawlers, not for assistants acting for one person. Treat a rule like this as a public statement of what you prefer, and enforce anything important at the edge.

The Personal Agent Protocol: what to prepare for

Meta's Meta business post on personal agents previews the Personal Agent Protocol, an open standard that Meta and Sierra are developing with Genesys, NiCE, Decagon, Rocket, Shopify, Stripe and Walmart. Meta says it will define how personal agents "securely convey user identity", how businesses can "steer personal agents for efficient website use or towards optimized APIs and skills", and how personal agents talk to business agents.

The protocol is not published yet, so there is nothing to install today. But the direction is clear, and you can prepare in ways that help with every agent, not just Muse.

• Decide your preferred paths. Meta's example is "use this flow for bookings, use that connector for order tracking, check this structured source for real-time inventory". Write those paths down now.
• Publish clear machine-readable facts. Product feeds, accurate structured data and an llms.txt file that points to your key pages all help an agent take the right route.
• Look at APIs you already have. If you offer an API for partners, it may become the faster path for agents too. Our guide to MCP and APIs for agents explains the options.

Three examples

These are illustrative scenarios based on the problems we see most often in site audits, not named client cases.

A family-run hotel

The hotel noticed short sessions from data-centre IPs on its rooms page, followed by abandoned bookings. The AI agent log checker showed muse- on most of them. The booking widget loaded room prices only after a slow third-party script and asked for a CAPTCHA on the dates step. The hotel moved the CAPTCHA to the final payment step and added the room rates to the page as text. Agent-led bookings started to complete, and the owner saw no rise in spam.

A sneaker shop

Limited releases attract bulk-buying bots of every kind. The shop kept Muse welcome on product and size pages, but added a Cloudflare rule that sends a managed challenge to verified agents on the checkout path during release windows only. Normal stock sells to agents and people alike; drops stay fair.

A local accounting firm

The firm worried that agents would read its client portal. They cannot: the portal needs each client's login and two-factor code, and Muse hands those steps back to its user. The firm left the portal unchanged, made sure its contact form worked without a CAPTCHA, and added its fees and services to the page in plain text so agents could answer questions like "how much does a small business tax return cost here?"

Common questions

Does Muse affect my Google rankings?

No. Muse is not a search engine and plays no part in Google or Bing rankings. What it affects is whether people who use it can do business with you.

Will Muse train Meta's models on my pages?

Meta's launch post talks about user data: people "can also opt out of their interactions being used to train Meta's AI models". It says nothing specific about the pages Muse reads for a task. If training is your concern, the control Meta documents for that is the Meta-ExternalAgent crawler, which you can block in robots.txt.

Is Muse traffic counted in my analytics?

Usually yes, because it runs a real browser that executes your analytics script. Expect short, direct sessions from US data-centre locations. If your analytics tool can segment by network, you can estimate how much of your traffic is agents.

Can I talk to Muse users directly?

Not through Muse itself. But a page that answers the common questions clearly, with prices, availability and next steps, is the message the agent will carry back to its user.

How does Muse compare with OpenAI dots and Grok Bot?

All three are always-on personal agents with their own cloud computers. They differ most in how they identify themselves. Read our guide to personal AI agents for a side-by-side table, and what OpenAI dots mean for your website for the OpenAI side.

Your Muse checklist

• □Search a week of logs for muse- with the AI agent log checker, and note which pages it visits.
• □Decide your position per area of the site: browse, book, buy, account. Write it down.
• □Remove CAPTCHAs from search, product and booking pages. Keep risk-based checks on account creation and high-value actions.
• □Put prices, stock and hours in plain HTML with matching structured data.
• □Offer guest checkout and at least one mainstream wallet.
• □If you need to block, use your CDN's verified bot identity for Muse, not a robots.txt line.
• □Leave Meta-ExternalTest allowed if you use Meta business integrations.
• □Re-test your key pages with the AI Agent Readiness Check after each change.

For the full list of agents and crawlers we track, see the AI and LLM bot directory.