Chroma MCP Server

by Chroma

Give Claude direct access to a vector database and suddenly every conversation has long-term memory. The Chroma MCP Server connects Claude Desktop (or any MCP-compatible client) to Chroma, the open-source embedding database, letting you store documents, run semantic searches, and manage collections without leaving your chat window. The standout capability here is semantic document search. Instead of relying on exact keyword matches, you can ask Claude to find documents by meaning. Store your engineering docs, support tickets, or research papers in a Chroma collection, then query them with natural language. Claude handles the embedding and retrieval behind the scenes using any of six supported embedding providers: the built-in default model, OpenAI, Cohere, Jina, Voyage AI, or Roboflow. Setup takes under five minutes for the basic ephemeral mode. Add one JSON block to your Claude Desktop config, restart, and you have 12 tools available immediately. For anything beyond experimentation, switch to persistent mode to keep your data across restarts, or point it at a self-hosted Chroma instance or Chroma Cloud for team-wide access. The server exposes fine-grained control over collections and documents. You can create collections with custom HNSW index parameters, add documents in bulk with metadata, run filtered queries combining semantic similarity with metadata constraints, and paginate through large result sets. Batch operations make it practical to index hundreds or thousands of documents in a single session. One thing worth noting: this is the official server built by the Chroma team (chroma-core on GitHub), not the community-built chroma-mcp-server package. It runs via uvx with zero dependency management on your part. Python 3.10+ is the only prerequisite. The server supports environment variables and dotenv files for credential management, so API keys for cloud deployments and embedding providers stay out of your command-line history.

FAQ