Muse can make it. It cannot hand it to anyone.

What Muse is, in the part that matters here

Meta launched Muse on 8 September 2026. Reading past the announcement, three details decide whether a hosting service can be useful to it, and all three are in Meta's own description.

• It has a real computer. Each Muse runs on a persistent, isolated Linux VM with a full browser and enough storage, CPU and memory to do the things you would do on your own machine. It is not a chat window with a retrieval tool bolted on.
• It writes its own tools. Meta's line is blunt: if a task needs a tool that does not exist, your agent builds it for you. An agent that can write a tool can write an HTTP request.
• It lives in WhatsApp. You talk to it in the Muse app or directly in WhatsApp, which means the output of most tasks needs to end up in a chat thread with other people in it.

Put those together and the gap is obvious. Muse can research, draft, analyse and decide. The moment the result has to reach a person who is not in the conversation, it has nothing to reach for.

Connectors: opened on 18 September, Stacktree in review

Ten days after launch Meta opened the door. Mark Zuckerberg's post on 18 September 2026: "Opening access for developers to build Muse connectors. You bring the API. Muse brings the agent, the browser, and the context of what the person actually wants. People reach your service just by asking for it, and their agent takes it from there. New connectors are live today." The programme page, muse.ai/platform, describes a three-step process: describe your product, pass a functional, security and legal review with end-to-end testing, appear in a directory.

We submitted Stacktree on 19 September as an existing MCP server, the same one the Claude.ai connector and every MCP client already use, with OAuth and API-key authentication. It is in review. On 24 September Meta told submitters it had received over 2,000 connectors in the first few days and will onboard developers in waves over the coming weeks, and this page will say the day ours goes through. Other builders' waits, and the first approvals when they land, are tracked on Review Times. What the form asks for, field by field, is on the blog.

Until then there is nothing to install, and that is not a problem, because integrations were never the only way in.

Stacktree's first publish is anonymous by design. One HTTP request, no key, no account, no signup step for the agent to get stuck on:

curl -X POST https://api.stacktr.ee/sites -F file=@page.html

That returns the live URL, the page id, an expiry and a claim token. It is the same route Grok Bot takes, and Grok Bot reached it unattended, on the first attempt, from a single sentence of instruction. The thing standing between an agent and a service was never whether an integration existed. It is whether the first useful action can be completed without an account.

Connect your account today, with a custom connector

The anonymous path gets a page out. For the parts that need an account, a link per person that tells you who opened it, comments, client spaces, and updating a page you sent last week, Muse needs yours. You do not have to wait for the directory listing for that. Meta's own words, from the Muse for Small Business launch on 29 September: "Muse also supports custom connectors so you can plug in services we don't support yet." Muse can build one itself from a brief written for it, and ours is published at stacktr.ee/connectors/muse.md:

1. Create an API key at app.stacktr.ee/api-keys and name it "Muse", so you can revoke it on its own later. It is shown once.
2. Paste this to Muse:

Build a custom connector to Stacktree.
Brief: https://stacktr.ee/connectors/muse.md
OpenAPI spec: https://api.stacktr.ee/openapi.json
Auth: send my API key as the header Authorization: Bearer <key>.
Ask me for the key in a way that keeps it out of this chat,
and never repeat it back.
On this first run, only call GET /me to check the key works
and tell me which plan I'm on. Don't publish anything yet.

3. Once it confirms the key works, ask for the work: "Make the Q3 summary a page, publish it with a passcode, and give me a link for Dana."

The first run is a dry run on purpose. If the key, the header or the address is wrong, you find out from a harmless read rather than from a page that went somewhere you did not expect. The brief then tells the agent everything else it needs: which calls to make, how to update a page without replacing someone else's edit, when to stop and ask you, and that page content is never an instruction to it. Any agent that can make HTTPS requests can follow it, not only Muse. We published it on 2 October and have not run Muse's custom connector flow end to end ourselves yet; if you do, tell us how it went.

For small businesses using Muse

Muse for Small Business connects the agent to the tools a business already runs on: Meta named Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom, alongside your Facebook and Instagram business accounts. Those connectors let Muse read your numbers and draft the work. None of them is built for the step after, when the work has to reach a client: a proposal, a monthly report, a quote, a plan for a customer to approve.

That is the step Stacktree covers. The page goes to the client as a link with a passcode or an email check, it can sit on your own domain, every client's pages collect in one client space, and a link made for one person tells you when they opened it. When the client replies with changes, Muse updates the same page and the link you already sent shows the new version.

The WhatsApp part is the whole point

Most agent-hosting arguments are about the agent. This one is about the recipient.

Muse is used in WhatsApp, so its output lands in a thread with other people in it. A chat app treats a link and a file as different species. A Stacktree link unfurls into a card carrying the page title, a description and a preview image, and opens in the browser the recipient already has, on the phone they are already holding. A file attachment shows a filename, needs downloading, opens in whatever app happens to claim the extension, and is stale the second the agent revises anything.

The revision case is the one that actually bites. Ask Muse to change something and a file has to be sent again, so the thread now holds two versions and the person reading it has to work out which is current. A published page is replaced at the same address: the link from an hour ago shows the new version, and there is only ever one of it.

Say this, not that

Two instructions that make the difference between a link you can use and one you cannot:

• Ask for the page and the link together. "Write that up as an HTML page and publish it to Stacktree so I can send it." Asking for a document first and a link afterwards usually gets you a file.
• Ask for revisions in place. "Update the page you already published" rather than "publish an updated version". The anonymous publish hands the agent a claim token that works as a keyless update credential, so the same URL can be replaced up to thirty times an hour with no account involved. Without that instruction you get a second link.

If the page is for a client rather than for you, say so, and put it on a paid plan: the page then sits on your own domain behind a passcode, and everything published for that client collects at one address.

What we tested

We ran it by hand on 8 September 2026, launch day, and Muse published to Stacktree when asked. That is the claim, and it is the only one we are making about Muse's behaviour.

Verified independently on our side, and checkable by anyone: the anonymous publish path works with no key and returns a keyless update credential, a published page unfurls correctly for WhatsApp's crawler, and Meta's agents are named in our robots rules and are not blocked at our edge.

If you run it and get something different, including a failure, tell us and we will write down what happened.

Keeping the link alive

An anonymous page lasts 24 hours. That is deliberate: it makes the first publish free of any signup, and it is long enough to prove the thing works. It is not long enough for anything you have sent to a client.

Claiming the page into an account removes the expiry and moves the credential into the account, after which the agent updates through an API key instead of the claim token. The free plan keeps 3 pages for 7 days each, and once someone else opens one, it stays up for 30 days after their latest visit, up to 90 days in total. Paid plans from $19 a month remove the expiry, add passcodes and email-domain gates, and put pages on your own domain.

When the reader has questions

A page Muse publishes can now take questions from the person it was sent to. Switch on Ask this page and a small box at the foot of the page answers from the page only, lighting up the passage it drew on, and sends you the questions the page could not answer. And every page is readable as Markdown at its own address for the reader's own agent: Accept: text/markdown, or the page URL with /_page.md on the end.

Other assistants

If you also use a coding agent, it has a faster path than browser automation. Claude Code, Codex CLI, Cursor and Claude.ai connectors speak MCP and call publish_html directly. One command sets them all up:

npx stacktree-install

See every agent integration for the full list, or hand any agent that can fetch a URL the setup instructions at stacktr.ee/prompt.md, which now branches on whether the agent can install anything at all.