MongoDB MCP Server is a Model Context Protocol server through which an AI client queries and manages MongoDB databases and MongoDB Atlas. It runs locally as a stdio or HTTP server against any self-managed or Atlas deployment given a connection string, or against the Atlas control plane given service-account credentials. MongoDB also hosts an Atlas-managed version that clients reach over OAuth, so nothing runs locally.
Database tools cover find, aggregate, count, insert, update, delete, explain, schema and index inspection, collection and database management, log access, and exporting query results as EJSON that the client reads back as a resource. Atlas tools create and inspect projects, clusters, database users, and IP access lists, load sample datasets, pause, resume, and scale clusters, read performance-advisor recommendations and alerts, and build, manage, and tear down Atlas Stream Processing resources. Atlas Local tools create and connect to local Atlas deployments, and assistant tools search MongoDB's own knowledge base of documentation and guidance.
Access is scoped by configuration rather than by trusting the model. A read-only mode registers only read, connect, and metadata tools, tools can be disabled by name, category, or operation type, an index-check mode rejects queries that would scan a collection, and destructive tools such as drop-database or delete-many request confirmation through MCP elicitation when the client supports it (aggregations whose stages write to a collection always do). Documents and bytes per query, count time, idle time, and active connections are capped, telemetry can be switched off, and an HTTP deployment can expose a separate monitoring endpoint for health checks and Prometheus metrics.
The server ships as plugins for Cursor, VS Code, Claude, Codex, Gemini CLI, GitHub Copilot CLI, and Grok Build, and configuration guides cover Devin, Claude Desktop, and other clients. The CLI is meant for single-user, loopback or private deployments and has no built-in authentication. For multi-tenant or public deployments the underlying library packages let you terminate authentication at your own proxy and key connection scope on a verified end-user identity, or you use the Atlas-managed server instead.
Features
- Database tools: find, aggregate, count, insert-many, update-many, delete-many, explain, collection schema and indexes, create, rename, and drop collections, database stats, and logs
- Atlas tools: projects, clusters (free, dedicated, upgrade, pause and resume), database users, IP access lists, sample datasets, performance advisor, alerts, and Stream Processing resources
- Atlas Local and assistant tools: create and connect to local Atlas deployments, and search the MongoDB Assistant knowledge base of documentation
- Exports as resources: query or aggregation results are exported in EJSON and read back through an exported-data resource, with automatic expiry and cleanup
- Read-only mode: registers only read, connect, and metadata tools so create, update, and delete operations are unavailable
- Tool filtering: disables tools by name, by category (atlas, mongodb), or by operation type (create, update, delete, read, metadata, connect)
- Confirmation via elicitation: destructive tools request user confirmation through the client, and aggregation stages that write to a collection always do
- Index check mode: rejects queries that would perform a collection scan
- Query limits: caps on documents and bytes per query, count time limits, idle timeouts, and maximum active connections
- Transports and deployment: stdio or HTTP transport, Docker images, an Azure deployment guide, and an optional monitoring server with health check and Prometheus metrics
- Hosted alternative: the Atlas-managed MCP server authenticates per user over OAuth and needs nothing run locally
- Agent plugins: bundled plugins for Cursor, VS Code, Claude, Codex, Gemini CLI, GitHub Copilot CLI, and Grok Build
