What you're building

Muse is Meta's personal AI agent. It reaches third-party services through connectors.

The important part, and the part that surprises people: there is no app to register and no app id to collect. Meta does not run your code. You expose an HTTP API, and Muse reads a description of it and calls it on the user's behalf.

Two files decide everything

Muse reads an OpenAPI document and an llms.txt from your service, then calls your API.

File	Who reads it	What it does
/openapi.json	the platform	Machine-readable: which operations exist, what they take, what they return
/llms.txt	the model	The same in prose, plus the rules — when to ask the person first, how to wait on slow work

That's the whole integration surface. There is no SDK to install and no callback to implement.

Muse ──(connector token)──▶ your connector ──(your credential)──▶ your API
        scoped, revocable      this template       held server-side

What the template gives you

Muse calls your API from Meta's infrastructure, acting for a person who is not sitting there watching. Three things follow from that, and they are what this template exists to handle:

1. The model must not hold your API key. Users get a token that names one connection. It is revocable, scoped, and carries no upstream secret. Your real credential stays server-side.
2. Slow, paid work needs wording, not just code. If the model retries a POST to "check progress", the person is charged twice. See Async & paid work.
3. The description is the interface. A model decides whether to call you based on prose you write. "Reads your balance" is weaker than "Read-only. This cannot move funds."

What you get out of it

The output is a plain, authenticated OpenAPI service. Muse reads it — but so can anything else that speaks OpenAPI: ChatGPT, Claude, Cursor, your own agent.

Not locked in

Nothing here is Muse-specific except some wording conventions in llms.txt, and those are yours to change. There is no account with us, no key from us, and no telemetry. Deploy it anywhere; delete the parts you don't want.

Next

• Quickstart — running locally in five minutes.
• Make it yours — the three files you actually edit.
• Want to know whether this is worth it? The Custom Connector path needs nothing from Meta, so you can validate before investing in anything.