Add execution receipts to a LangChain agent: check what it ran, not what it says it ran.

Route 1: the langchain-zambo package

The packaged route turns Zambo's MCP server into LangChain tools in one import. langchain-zambo (on PyPI, version 0.1.4) is a thin wrapper over langchain-mcp-adapters: it converts the tools served at https://zambo.dev/mcp into LangChain StructuredTool objects at runtime, over streamable HTTP, with no auth and no API key. Nothing is reimplemented in the package; the tools your agent receives are the server's own.

import asyncio
from langchain_zambo import get_zambo_tools

async def main():
    tools = await get_zambo_tools()  # Zambo MCP tools as LangChain tools, no API key
    prices = [t for t in tools if "price" in t.name]
    print("price tools:", [t.name for t in prices][:5])

asyncio.run(main())

From there the tools drop into a LangGraph agent like any others. The difference shows up after a call: every call returns its result plus a verifiable receipt, the record of that call captured by the server at execution time and published at a public URL. If you already run a MultiServerMCPClient, server_config() returns the Zambo server entry so you can merge it with your own servers instead of replacing your setup.

Route 2: the direct MCP pattern

The package is a convenience, not a requirement. Any LangChain or LangGraph agent that can use MCP tools can point langchain-mcp-adapters straight at https://zambo.dev/mcp and get the same behavior: calls execute on the server, and each one leaves a receipt because the executing layer makes the record. That is the general pattern behind MCP server receipts, and it works the same from any framework.

One habit completes the pattern: end multi-step jobs with workflow_close. The run's step receipts bind into a summary receipt with the steps as children and a single root digest over them, so the job you hand to a client, a reviewer, or your future self verifies as one sealed record, not a folder of separate claims.

Checking the run afterwards

• Per call: open the receipt URL and recompute the commitment from its canonical bytes, by hand or with the live verifier. A match means the saved record is the record made when the call ran.
• Per run: recompute the workflow root from the step receipt IDs and compare it with the summary receipt's root. A missing, added, or swapped step breaks it.
• As an audit: the receipts are the audit trail, countable and checkable after the fact. The method is written up in how to audit AI agent tool calls, and the verifier walkthrough runs the procedure against a real public receipt.
• Full days covered: a Day Pass ($0.99) keeps every call receipted for 24 hours, and sealed runs keep their page, verify badge, and PDF pack after the free viewing window.