Chapter 21
MongoDB on Claude Managed Agents
MongoDB on Claude Managed Agents
Setup boilerplate for the Fraud Review Agent with MongoDB Atlas and Claude Managed Agents
cookbook. The teaching code lives inline in the notebook — the four retrieval pipeline
builders, the custom-tool handlers, and the requires_action gate loop are all defined there.
This package holds only the pieces the notebook imports rather than reads.
| Module | What's inside |
|---|---|
config.py | Tunables, index names, and the MongoDB server-version check (supports_rank_fusion). |
embeddings.py | An embedding + rerank client that adapts the MongoDB Atlas AI endpoint to the voyageai interface, so the cookbook is provider-agnostic. |
tools.py | MongoDB Atlas setup — seed the collection, create the vector + Atlas Search indexes and wait until they are queryable and synced, preflight — plus the shared decision/audit document shapers used by the notebook's record_decision handler and the AP2 module. |
ap2_mandates.py | AP2 (Agent Payments Protocol) mandate signing and verification (ES256 JWTs) — a crypto black box the notebook calls through verify_mandates and acts on its verdict. |
seed.py | Loads the plaintext fixture from ../example_data/mongodb_on_cma/. |
The MongoDB credential (MONGO_URI) only ever lives on your side of the boundary: pymongo
runs in the notebook's host-side handlers, never in the agent context or its sandbox. See the
cookbook for the connection pattern and the end-to-end fraud-review agent.