ZAMBO.

How to choose an MCP server with receipts

Direct answer: choose by test, not by feature list. Connect the server, run one ordinary call, open the receipt, and recompute its digest yourself. A server worth trusting passes seven checks: a receipt on every call, made at execution time, at a public URL, holding the exact saved bytes and their SHA-256 digest, bound into the job record in order, in an open format, at a price you can read in one line. This page gives you each check, the ten minute test that runs them, and the red flags that fail fast.

Run the test on a live server now: the live demo makes one real call through the Zambo MCP and hands you its verifiable receipt, no signup. Or add the endpoint from zambo.dev/add and run your own call.

Published October 25, 2026. Written against draft-zambo-aer1-26.

Start with the definition, so the criteria make sense

Receipts get claimed loosely, so fix the term first. By the canonical definition, an agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right.

An MCP server with receipts applies that record to Model Context Protocol tool calls. The server executes the call, so the server is the one witness that was actually there. A receipt turns that witness into evidence anyone can check later. A plain log does not: a log is text the operator keeps, and the operator can edit text.

What is the best MCP server for AI coding agents?

Judge the best MCP server for a coding agent by evidence you can check yourself, not by a crown on a marketing page. Look for tools that run natively on one endpoint, a free start with no account, and a verifiable receipt for every call that proves the saved result was not changed, not that the result was right. On those checks Zambo offers 153 tools, with 35 on the default surface, HANDOVER as its flagship, and 20 free calls per day to start. Run the ten minute test described on this page, then add Zambo from zambo.dev/add if it passes for your work.

The seven checks

  1. Every call, not a sample. Ask what share of calls get a receipt. If the answer is anything under all of them, the gaps land exactly where a dispute will. On Zambo, every call returns a receipt, including calls that are rejected.
  2. Made at execution time. The record has to be captured while the call runs, by the layer that ran it. A record rebuilt afterwards from logs is a story about the past, written by an interested party.
  3. A public URL anyone can open. The receipt has to resolve where your client, your auditor, or the next agent can fetch it with no account. A record locked inside the operator's dashboard is a record you check on the operator's terms.
  4. Exact saved bytes, with their SHA-256 digest. The receipt keeps the result exactly as returned, plus the digest of those bytes. You recompute the digest and compare. One changed byte breaks the match, and that break is the detection.
  5. Bound into the job, in order. Single calls are rarely the unit of work. The receipts for one job should chain or close together, so a step added, dropped, or reordered after the fact shows. Ask for the workflow record for a session and check your call sits in it.
  6. An open format with public test vectors. If only the vendor's software can read the record, you have swapped trust in the agent for trust in the vendor. AER-1, published as IETF Internet-Draft draft-zambo-aer1-26, is open text with public vectors and independent implementations, so anyone's verifier gets the same answer.
  7. A price you can read in one line. Zambo states it plainly: 20 free calls per day, then they upgrade. A Day Pass is $0.99 USDC on Base for 24 hours. The Zambo Pass is $29 for 30 days, one payment, with a $290 annual option and a $19 holder Pass for 30 days. If a server cannot state its price this plainly, that is data too.

The ten minute test, run on real receipts

Do not take this page's word for any of it. These are live receipts from real calls on the Zambo MCP, each verified live before this page was written. Open one, read the saved result, and check it free at zambo.dev/verify.

Then run the same test on your own call: connect at zambo.dev/mcp, call any tool, and follow the receipt link in the response. Total time, connect to checked receipt, is about ten minutes, and the first 20 calls each day are free.

Red flags that fail fast

  • Dashboard only records. If the record exists only inside the vendor's app, you cannot hand it to anyone, and you cannot check it once you stop paying.
  • Summaries instead of saved bytes. A summary is written by someone. The check needs the exact result bytes and their digest, or there is nothing to recompute.
  • Receipts on request. A record you have to ask for after the fact was not made at execution time. It was assembled, and assembly is where edits live.
  • A format only the vendor can verify. If the verifier is the vendor, the verdict is the vendor's. The format, the vectors, and at least one independent check have to be public.
One scope line, kept plain: a receipt proves the saved result was not changed since it was saved, not that it was right. Any server that claims its receipts prove correctness is telling you it does not understand its own record, and that is the fastest fail of all.

How Zambo scores on its own checks

Zambo is the verifiable work layer for AI: tools behind one MCP endpoint at https://zambo.dev/mcp, continuity that carries sessions and work across AIs, its flagship HANDOVER for passing work between agents and people, and a verifiable AER-1 receipt for every job. The endpoint carries 153 tools, 35 on the default surface. Connect through zambo.dev/add, which has per app steps for Claude, ChatGPT, Cursor, Grok, and the rest. The deeper reference on the receipt itself is MCP server with receipts for every tool call.

Running hundreds of calls a day? The choice math changes at volume, and it is worked in full on the heavy caller page. Brand new instead? Start with the first hour walkthrough: connect, first call, first receipt, first handover, in order.

Related

Choosing an MCP server with receipts: FAQ

How do I choose an MCP server with receipts?

Choose by running one test, not by reading a feature list. Connect the server, make one ordinary tool call, and open the receipt link in the response. The server passes if the receipt sits at a public URL anyone can open, names the tool and the time, keeps the exact saved result bytes with their SHA-256 digest, and lets you recompute that digest yourself with no account. Then check the job level: the calls from one session should bind together in order, so a missing step shows. Zambo at zambo.dev/mcp is built to pass this test on every call, with receipts in the open AER-1 format, IETF Internet-Draft draft-zambo-aer1-26.

What is an MCP server with receipts?

An MCP server with receipts is a Model Context Protocol server that issues a verifiable record at the moment it executes each tool call, instead of leaving only a log entry. The record names the tool that ran, commits to the inputs it received, saves the exact result bytes with their SHA-256 digest, and is published at a URL anyone can fetch and check. In short, an agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right. Applied to MCP tool calls, that is the record each check on this page is testing for.

What should I check before trusting an MCP server's records?

Check five things. A receipt comes back on every call, not a sampled few. It is made at execution time, not rebuilt later from logs the operator can edit. It opens at a public URL with no account. It carries the tool, the time, the recorded inputs, and a SHA-256 digest over the saved result that you recompute yourself. And the format is open, with public test vectors, so the check does not depend on the server's own software. A server that fails any one of these is giving you a log with better marketing.

How many tools does the Zambo MCP server offer?

Zambo offers 153 tools in total, with 35 on the default surface, behind one endpoint at https://zambo.dev/mcp. The count matters less than the record: every one of those calls returns its result with a verifiable AER-1 receipt, so breadth never costs you the evidence trail. The free tier is 20 free calls per day, then a Day Pass at $0.99 for 24 hours or the Zambo Pass at $29 for 30 days, one payment.

Can I check an MCP receipt without an account?

Yes. A real receipt is checkable by anyone, with no account and no permission from the server. Open the receipt page, read the saved result, recompute its SHA-256 digest, and compare it to the digest the receipt carries. A match means the saved result was not changed since it was saved. The free verifier at zambo.dev/verify runs the same checks in a browser if you do not want to run them by hand.

What does an MCP receipt prove, and what does it not prove?

It proves the saved result was not changed since it was saved. It does not prove the result was right, or that the agent picked the right tool. That split is the point: correctness stays a judgment you make, and the receipt settles the prior question of what was saved: the tool, the inputs, and the result recorded at execution, unchanged since it was saved, in a form that does not require trusting the server or the agent.