MongoDB MCP Server

MongoDB

MongoDB stores data in databases → collections → documents. Documents are JSON-like with _id. Works with Atlas, self-hosted, DocumentDB, Cosmos DB MongoDB API. Connection string format: mongodb+srv:// or mongodb://.

Operational order (unfamiliar database)

1. List databases (when URI allows) and list collections — learn collection names.
2. Inspect collection or get AI description — sample field shapes and indexes before filtering.
3. Get documents — pass collection always; use query to list (with skip/limit) or document_id for one _id.
4. Aggregate when you need $group, $lookup, or analytics — one collection, pipeline as JSON array.
5. Export / import — export collection to JSON or JSONL; large exports return a presigned download_url. Re-import with mongodb_import_documents via file_url (export link), file_content_base64 (≤1 MB), or local_path with MCPBundles Desktop for files on this computer.
6. Explain slow finds or pipelines before wide scans.
7. Indexes — list indexes, create, or drop named indexes; create/drop collections only when the user intends schema changes.
8. Insert / update / delete / replace only after you know the collection schema; update and delete filters cannot be empty {}.

Data Model

• Databases — namespaces that contain one or more collections.
• Collections — document containers; every read or write targets a collection name.
• Documents — BSON keyed by _id; JSON responses expose ObjectIds as _id.$oid or string hex in descriptions.
• Queries — filter JSON on get documents, search collections, update, and delete (Extended JSON $oid supported).

Two credential field setups

Gotchas

• Connection string: Username and password can live inside the URI — separate username/password fields are optional when the URI is set.
• Single document by id: pass document_id on get documents (plain hex, Extended JSON $oid, not colon handles).
• Cross-collection discovery: use search collections with a non-empty filter; prefer get documents when the collection is known.
• Destructive filters: update documents and delete documents reject {} filters.
• Large exports: mongodb_export_collection returns inline JSON/JSONL under ~512 KB; larger payloads spill to S3 with download_url — pass that URL to mongodb_import_documents as file_url.
• Large imports: use file_url or Desktop local_path when the file exceeds the 1 MB inline base64 limit.
• Local MongoDB: Use the Local MongoDB on this computer setup — not a connection string with localhost in it.
• Atlas connection: Use mongodb+srv:// with credentials in the URI. Network access must allow MCPBundles egress.