Your Muse agent did the work. Here's how it proves it.

The Muse version of the problem

A personal agent is not a chatbot you watch. You brief it in the morning, it works while you are in meetings, on your phone, away from the chat entirely, and later it reports: handled. The report is the product surface, and the report is written by the same process that did the work. Nothing inside it was ever compared against anything outside the model that wrote it. A fabricated finish and a real finish arrive in the same calm voice, and the more capable the agent gets, the longer and more detailed its unsupervised runs become.

Muse owners meet this in its purest form, because a Muse agent works across apps and connectors, on jobs that span hours. The fix is not a better summary. The fix is a record underneath the summary, saved at execution by a system that is not the agent, checkable by the owner or by any other AI the owner trusts more that day.

Add Zambo to Muse: the Muse connector route that works today

The skill checks records. Zambo is where the records come from. Muse connects to remote MCP servers as custom connectors the agent builds itself, and Zambo is a remote MCP server with no API key on its free tier. The whole setup is one message in any Muse chat:

Create a custom connector for Zambo, a remote MCP server.
Name: Zambo
URL: https://zambo.dev/mcp
Transport: remote streamable HTTP
Auth: none, no API key
Connect to it, test the tools, and save it so you can use Zambo in future conversations.

Muse builds the connector, connects, and tests the tools itself. There is no settings form for adding a server, nothing to install, and no key to paste. Once connected, you can review or disconnect the connector any time under Settings, then Connectors. The full walkthrough lives on the companion page, Use Zambo with Muse.

One honest note, stated plainly: Zambo's submission to the Muse connector directory is in review. Meta confirmed receipt of the submission on September 23, 2026, and the directory review is still open. So this page claims no directory listing and no partnership, because neither exists today. The custom connector above is the working route: the same server a directory listing would point at, the same receipts, available to any Muse account right now.

A worked example, run for real

On October 21, 2026, we ran a small morning-brief job through the Zambo MCP endpoint, the same calls a connected Muse agent makes, under one session. Three steps, three receipts, then the package a reviewer signs off on.

The third step is the one that teaches the lesson. An unchecked summary of that job could have read "found roofing leads in Tampa," and nothing in the prose would have betrayed it. The record says the search ran and found none. That is the entire argument for checking records instead of summaries, demonstrated by a real run instead of asserted by a vendor.

The three steps were then packaged as a handover, one link a reviewer opens to walk the steps and record a decision: zambo.dev/handover/ea028837, closing receipt 4ead905f.

Then the skill's step 2, run live against all four receipts: the verify API recomputed the digest over each saved result and compared it with the hash recorded when the call ran. For the price receipt, recomputed sha256:40f6ec0d2626cfcebeb020b4e0204fd1826a0d5674a84887e4cc4d2a719188ed against recorded sha256:40f6ec0d2626cfcebeb020b4e0204fd1826a0d5674a84887e4cc4d2a719188ed: a match, status verified. The scan, the search, and the handover closing receipt recomputed the same way, all verified. Run any of them yourself from the receipt links above, or paste an ID into zambo.dev/verify. The independent proof stays independent: the receipt at the top of this page was left by a team with no connection to Zambo, testing the format on their own, and it re-verified again today.

What the receipt does not prove

A verified receipt proves one narrow, load-bearing fact: the saved result has not changed since it was saved. It does not prove the result was correct, and it does not prove an outside event happened. Whether BTC was the right thing to check, whether blocking that email was the right call, whether a lead search with zero results was a good search: those judgements stay with the owner, now exercised against a fixed record instead of a confident paragraph.

And some Muse work leaves no receipt at all, because no tool call ran behind it. The skill's branch for that case is blunt: if step 1 comes back empty, there is nothing to check. The three honest moves are to rerun the job on a recorded path and check that run, to accept the claim knowingly as a decision where the stakes allow it, or to change the default upstream so the next job runs recorded. No record, no sign-off.

Frequently asked questions