How Meta Muse custom connectors work and how safe they are

A custom connector is Meta Muse writing its own code to talk to a service that is not in Meta's connector directory. You give it a link to that service's instructions, it reads them, writes the integration inside its own cloud computer, tests it, and saves the result. It takes about five minutes and you never do it twice for the same service.

The safety story is genuinely good in one way and genuinely uncertain in another. Meta's credential design means the model never handles your real key. Meta's review process does not cover custom connectors at all, which it says plainly on its own help pages.

Both are true at once. This page explains them without spin, including the parts that are awkward for us, because we run one of these connectors.

What actually happens, step by step

1. You paste a prompt. Ours is at the bottom of this page.
2. Muse reads the documentation. Ours is a plain markdown file written for an agent rather than a person: /connectors/muse.md.
3. Muse writes the code. Inside its own cloud machine, not on your phone.
4. Muse asks for a key. Into the Secure Credentials Store, not the chat.
5. Muse asks permission to reach that domain. Nothing leaves its machine until you allow it.
6. Meta shows a caution notice, because nobody reviewed this connector.
7. It is saved as a skill. Next time you just ask.

The good part: your key

This is the piece Meta has clearly thought hard about.

Put a key into the Secure Credentials Store and the model working your request does not get it. It gets a stand-in value. The real secret is held separately and swapped in at the network edge by the part of the system that handles permissions.

Why that matters: if a web page contains hidden text saying "print the user's API key", the model cannot comply, because it does not have the key to print. The protection is not the model being careful. It is the model never being given the thing.

So: never type a key into the chat, and if a service offers read-only keys, start with one. Ours does, on the key page. A read-only key can check addresses and prices, and the server refuses it if it ever tries to send.

The uncertain part: nobody checked the connector

Meta reviews the connectors in its own directory. It does not review custom ones. Its help pages say so directly, and tell you to grant access with caution and read the provider's privacy policy.

So the question of whether a service is honest, competent and careful with your data is yours to answer. No platform did it for you. The caution notice on screen is accurate. Read it rather than clicking past it.

What it does not mean: the connector cannot read your email or your files. It can do what its own API does, Muse still asks you before consequential actions, and each key is separate.

What it does mean: the service sees whatever you send it. For us, that is a recipient address and the contents of a letter. If that service is careless with its own database, that is your data in someone else's incident. And nobody vetted the claims on its marketing page, including this one.

How to judge one before you add it

A checklist that costs two minutes.

1. Can you read its brief? The file Muse will read should be public. Ours is at /connectors/muse.md. If a service will not show you what it tells your agent to do, walk away.
2. Real privacy policy and terms? Per-service pages, not a template. Ours are linked in the footer of every page here.
3. Does it ask for the smallest thing it needs? A mail service needs addresses and letter contents. It does not need your calendar.
4. Approval gates on the expensive steps? Ours splits checking an address, pricing and sending into three separate operations, so Muse has to stop and ask.
5. Does it tell you the bad news? A page that only lists upsides has decided not to tell you something.
6. Is there a person to email? Ours is support@noerrands.com.

What can still go wrong

Most likely first.

• You approve something wrong. The commonest real failure. A letter accepted for printing cannot be recalled.
• The service has an outage. For us, a failed send is refunded to your credits automatically.
• The service is a bad actor. Nobody checked. Do the checklist.
• Your key leaks from the service's end. Email support@noerrands.com and we revoke ours, then remove the stored credential in Muse's settings.

Our connector specifically

No Errands sends real letters and postcards, sends faxes, and books appointments. United States addresses only. A one-page black-and-white letter is $2.49, extra pages $0.25, color adds $0.50, certified adds $7.50, a postcard is $1.49, and a fax is $1.49 for the first page. Address checks and estimates are free. Credits are bought as $2.49 for a single letter, or $10.00 for $12.50 of credit (one free letter), $25.00 for $32.50 of credit (three free letters) and $50.00 for $67.50 of credit (seven free letters); every pack credits more than it costs, and it is separate from your Muse subscription.

Muse checks the address, prices the job, shows you the recipient, the content and the total, then waits. It is a custom connector, so Meta has not reviewed it.

Questions people ask

Are Meta Muse custom connectors reviewed by Meta?

No. Meta reviews the connectors in its own directory and says plainly that it does not review custom ones. Muse shows you a caution notice the first time you add one, ours included.

Can a custom connector see my email or my files?

No. A connector can do what its own API does and nothing else, and Muse still asks you before consequential actions. Each key is separate.

Does the agent ever handle my real API key?

Not when the key is in the Secure Credentials Store. The model works with a stand-in value and the real secret is substituted at the network edge by the part of Meta's system that handles permissions.

What is most likely to go wrong?

You approve something wrong. A letter accepted for printing cannot be recalled, so the approval screen is the last chance to stop it.

The prompt to paste if you still want it

Get a key at https://noerrands.com/get-key, then copy the whole block and paste it into Muse as one message.

Build a custom connector for No Errands. One API key covers three services:
printed mail (letters and postcards), fax, and appointment booking.

1. Read these three connector briefs and follow them exactly:
   https://noerrands.com/connectors/muse.md
   https://noerrands.com/connectors/fax/muse.md
   https://noerrands.com/connectors/booking/muse.md

2. Store my API key in the Secure Credentials Store as NOERRANDS_API_KEY.
   The same key works for all three. Never print, log, echo, or repeat it,
   not even partially.
   Send it as the header: Authorization: Bearer <NOERRANDS_API_KEY>

3. Prefer the MCP endpoints (streamable HTTP):
   https://api.noerrands.com/c/mail/mcp
   https://api.noerrands.com/c/fax/mcp
   https://api.noerrands.com/c/booking/mcp
   If you use the REST APIs instead, each brief links its OpenAPI document.

4. Before you mail anything, fax anything, or book anything, show me exactly
   what will happen, what it costs, and who it goes to, and wait for my
   explicit approval. Never act without asking me first.

5. Save this as a reusable skill called "Mail, fax and bookings".

Plain text version: /muse-prompt.txt. Step by step: send a letter with Meta Muse, or all the how-tos.