Chroma MCP Server vs Qdrant MCP Server

When to pick Chroma MCP Server

• Your agent needs to maintain collections as well as search them. It can create collections, inspect their contents, rename them, change metadata, and delete them through dedicated tools.
• You need to correct or remove stored documents. chroma_update_documents changes content, metadata, or embeddings; chroma_delete_documents removes specific documents. Custom IDs also let you identify documents when adding them.
• Retrieval requires more than a natural-language similarity query. Chroma supports metadata and document-content filters, full text search, and retrieval by ID through chroma_get_documents.
• You are testing a retrieval workflow before choosing its database deployment. Start with an ephemeral, in-memory client, use a persistent directory for file storage, or configure the same MCP server for self-hosted HTTP or Chroma Cloud access.

When to pick Qdrant MCP Server

• Your workflow centers on saving information and recalling it later. qdrant-store accepts text and optional metadata; qdrant-find retrieves information relevant to a natural-language query.
• You want a searchable collection of reusable code snippets. The documented pattern stores a description in information and the code in metadata.code. TOOL_STORE_DESCRIPTION and TOOL_FIND_DESCRIPTION tell the model how to use that collection.
• You need a remotely accessible MCP process. Qdrant offers SSE and Streamable HTTP alongside its default stdio transport, with environment variables for the listening host and port.
• Your agent should search existing memories without adding new ones. Set QDRANT_READ_ONLY=true to disable qdrant-store. You can also use QDRANT_SEARCH_LIMIT to cap the results returned by searches.

Main differences

• Collection control differs. Chroma exposes collection inspection and document maintenance tools. Qdrant concentrates on storing and finding memories, automatically creating the named collection if it does not exist. Setting COLLECTION_NAME fixes the collection; without a default, tool calls must supply its name.
• Database location and MCP transport are separate choices. Chroma connects to local storage or an HTTP/cloud database while communicating with its MCP client over stdio. Qdrant supports a local database path or server URL and additionally offers network transports for the MCP connection itself.
• Search inputs differ. Chroma exposes IDs, metadata filters, and document-content filtering. Qdrant's documented find input accepts a query and, when needed, a collection name; its store tool accepts optional metadata alongside the text.
• Embedding configuration differs. Chroma lets you select an embedding function when creating a collection and persists that choice for subsequent inserts and queries. Its options include external API providers such as OpenAI and Cohere. Qdrant supports FastEmbed models, selected through EMBEDDING_MODEL.
• Credentials follow the database and embedding choices. Chroma Cloud uses tenant, database, and API-key settings; its HTTP configuration includes custom authentication credentials. External embedding functions need their provider keys too. Qdrant uses QDRANT_API_KEY for authenticated server connections, including Qdrant Cloud.

Setup

For Chroma, run uvx chroma-mcp for an ephemeral client. For file storage, use uvx chroma-mcp --client-type persistent --data-dir /full/path/to/your/data/directory. Configure cloud or HTTP access through client flags or environment variables. --dotenv-path selects a configuration file; command-line arguments override environment variables.

For Qdrant, connect to a local server with QDRANT_URL="http://localhost:6333" COLLECTION_NAME="my-collection" uvx mcp-server-qdrant. For file-based local mode, replace QDRANT_URL with QDRANT_LOCAL_PATH; do not set both. Add --transport sse or --transport streamable-http when configuring a network MCP connection.