How to Add a Custom Connector or MCP Server to Meta Muse

Muse ships with dozens of built-in Connectors (Gmail, Notion, Slack, Shopify and more), plus Zapier. But sooner or later you will want it to use a tool Meta has not built a Connector for. That is what custom Connectors are for, and they work with both regular APIs and MCP servers (Model Context Protocol, the open standard many AI tools use to plug into outside services).

The surprising part: there is no "Add MCP server" form. You set it up by talking to Muse.

Before you start: what you need

1. The address. For an MCP server, its public URL (for example, a hosted https endpoint). For a regular API, a link to its documentation.
2. How it authenticates. Usually one of: no auth, an API key sent in a header, or a bearer token. Check the service's docs for the exact header name.
3. The key itself, ready to paste into Muse's secure prompt (not into the chat).
4. What you want to do with it. "Look up order status" or "transcribe a YouTube video" helps Muse test the right things.

What works and what does not

• Works best: remote MCP servers over HTTP (streamable HTTP). Muse runs on a cloud computer, so a hosted server is the natural fit.
• Can work: stdio MCP servers that Muse can install on its VM with npm or pip.
• Will not work: anything on your own laptop or home network. Muse cannot reach localhost.
• Fiddly: services that only support OAuth sign-in. Muse may need its browser and a few extra rounds.

Step 1: Send Muse one clear setup message

Put everything in a single message. This template works well:

I want to add a custom connector.
Service: [name]
MCP server URL: [https://...]
Transport: remote HTTP
Auth: API key sent in the header "[header-name]"
What I'll use it for: [one sentence]
Please build the integration, ask me for the key using the secure credential prompt (not in chat), test each tool, and save it as a skill I can use in any conversation.

Muse may ask a couple of follow-ups, such as whether the server runs over HTTP or stdio, or which header carries the key.

Step 2: Enter the key in the secure prompt

When Muse needs your API key, it opens a secure entry flow. Paste the key there. Keys entered this way go into Muse's Secure Credentials Store, and Sentinel (Muse's separate safety agent) inserts them into requests at the network boundary, so the agent itself never handles the raw key.

If you paste a key into the chat by mistake, rotate it with the provider. It is now in your conversation history.

Step 3: Test every tool

Do not assume it works because Muse says "connected." Ask:

List every tool this connector has. Call each one once with a simple test and show me the raw result.

If a tool fails, the error usually points to a wrong header name, an expired key, or a URL typo.

Step 4: Set approval rules

Go to Settings > Connectors and set how much freedom this Connector gets. For a read-only data source, automatic is fine. For anything that writes, sends, posts or spends, keep approval on.

Step 5: Confirm it was saved

Start a new conversation and use it without re-explaining:

Use my [service] connector to [do the thing].

If Muse does not recognize it, ask it to save the integration as a skill again.

Worked example: give Muse YouTube transcripts

Here is a real example using a hosted MCP server. We run one at YouTubeTranscript.dev, so we know its setup details. It lets Muse fetch the full transcript of any YouTube video as text, which is handy for summaries, notes and research (more on that in How to use Meta Muse with YouTube videos).

What you need: an API key from your YouTubeTranscript.dev account (MCP access uses your regular API credits, on paid plans).

I want to add a custom connector.
Service: YouTubeTranscript.dev
MCP server URL: https://mcp.youtubetranscript.dev
Transport: remote HTTP
Auth: API key sent in the header "x-api-token"
What I'll use it for: getting transcripts of YouTube videos so you can summarize them.
Please ask me for the key using the secure credential prompt, test the tools, and save it as a skill.

The server exposes five tools: transcribe_v2 (get a transcript from a YouTube URL or video ID), get_transcript (fetch one you already made), list_transcripts (search your history), get_stats (your remaining credits) and delete_transcript. Note that transcribe_v2 reads the video's captions, so a video with no captions at all will not return text. Then test it:

Use my YouTubeTranscript connector to get the transcript of https://www.youtube.com/watch?v=VIDEO_ID and give me a 5-bullet summary.

If you only need a transcript now and then, you do not need a connector at all: our free transcript tool works without an account, and you can paste the text into Muse.

Safety checklist for custom Connectors

• Meta does not review them. You are trusting the service operator with whatever Muse sends. Read its privacy policy.
• Less isolation than built-in Connectors. Custom code does not get the extra privilege separation Meta applies to its own Connectors. Sentinel still mediates outbound traffic.
• Use scoped keys. If the service lets you create a key with limited permissions or a spending cap, do that.
• Keep keys out of chat. Always use the secure prompt.
• Same usage meter. Calls through custom Connectors count against your Muse allowance like everything else.

Prefer something more point-and-click? See whether the app is available through Zapier in our Muse Connectors guide.