Menu

AI

Chroma MCP Server: Tools, Setup and Security

Chroma's open-source MCP server lets AI apps create collections and run vector, full-text and metadata search. See its 12 tools, client modes and safety tips.

By: Irina ShvayaNovember 2, 2026

Chroma MCP Server: Tools, Setup and Security

This profile is part of our MCP server directory, within The MCP Guide. Profiles are checked against each vendor's own documentation. Need a server built? See MCP development.

The Chroma MCP server lets an AI app store and search text in Chroma, the open-source embedding database. Your assistant can create collections, add documents and find them again with vector search, full-text search and metadata filters. You run it locally with uvx, and it can point at in-memory, on-disk, self-hosted or Chroma Cloud storage.

Checked against: the chroma-core/chroma-mcp README, Chroma's MCP integration docs and the chroma-mcp package on PyPI (version 0.2.6), checked on 2 November 2026. eSEOspace did not install or test this server.

Key Takeaways

  • The server exposes 12 tools for managing collections and documents, including delete tools.
  • Four client types are supported: ephemeral (in memory), persistent (local files), HTTP (your own Chroma server) and cloud (Chroma Cloud).
  • The default is the ephemeral client, so data disappears when the server restarts unless you choose persistent or cloud.
  • There is no hosted Chroma MCP URL in the docs we opened; you run the server yourself.
  • The latest GitHub release is v0.2.6 from August 2025, so check the repo before you depend on it.

At a glance

ItemDetail
PublisherChroma (chroma-core)
Official?Yes
HostingLocal process via uvx chroma-mcp
Transportstdio (the documented configs all launch a local command)
Sign-inNone locally; API key for Chroma Cloud; custom auth credentials for self-hosted
Read or writeRead-write, including delete tools
PriceFree and open source (Apache 2.0); Chroma Cloud is billed by Chroma
Source codeOpen, Python
Statusv0.2.6 (released 14 August 2025)

What it can do

The README lists these tools:

ToolWhat it does
chroma_list_collectionsList collections, with pagination
chroma_create_collectionCreate a collection, with optional HNSW settings
chroma_peek_collectionShow a sample of documents
chroma_get_collection_infoDetails about a collection
chroma_get_collection_countNumber of documents
chroma_modify_collectionRename or change metadata
chroma_delete_collectionDelete a collection
chroma_add_documentsAdd documents with metadata and IDs
chroma_query_documentsSemantic search with filters
chroma_get_documentsFetch by ID or filter
chroma_update_documentsUpdate content, metadata or embeddings
chroma_delete_documentsDelete documents

Chroma's docs suggest uses such as a shared team knowledge base and adding memory to a chat. Supported embedding functions are default, cohere, openai, jina, voyageai and roboflow. A collection keeps the embedding function it was created with, a feature Chroma says arrived in Chroma v1.0.0.

If you want an assistant to search open-source package code instead, the README points to a separate Chroma product called Package Search with its own repository.

How to connect it

The README gives claude_desktop_config.json entries. The simplest one starts an in-memory client:

"chroma": {
    "command": "uvx",
    "args": [
        "chroma-mcp"
    ]
}

To keep data on disk, add the persistent client type and a folder:

"chroma": {
    "command": "uvx",
    "args": [
        "chroma-mcp",
        "--client-type",
        "persistent",
        "--data-dir",
        "/full/path/to/your/data/directory"
    ]
}

For Chroma Cloud, use --client-type cloud with --tenant, --database and --api-key. For your own Chroma server, use --client-type http with --host, --port, --custom-auth-credentials and --ssl. These snippets go inside your client's mcpServers object. Our MCP config generator can wrap them for Cursor, VS Code or Claude Code.

You can also set everything with environment variables such as CHROMA_CLIENT_TYPE, CHROMA_TENANT, CHROMA_DATABASE and CHROMA_API_KEY. The server reads a .chroma_env file in the working directory by default, or a path you pass with --dotenv-path. Command-line arguments win over environment variables.

Security notes

The tool list includes chroma_delete_collection and chroma_delete_documents. Keep tool approval on in your client, and consider blocking those two tools if the assistant only needs search. Our tool permission planner can help you decide.

Chroma's README says putting API keys in arguments "is fine on local devices," but suggests a .env file via --dotenv-path for safety. Embedding providers that call outside APIs need their own keys, named like CHROMA_COHERE_API_KEY. Keep those files out of shared folders and version control.

Documents you store come back into the model's context later. Text from customers or the web can carry hidden instructions, so read prompt injection through tool results before you index untrusted content. The MCP Security Center covers the wider risks.

Alternatives

Next step

Try the persistent client on a small, non-sensitive set of documents, then decide whether you need Chroma Cloud. If you are weighing retrieval against live tool calls, read MCP vs RAG.

More in the MCP Guide

Free tools: MCP Server Risk Checklist · Tool Permission Planner

Latest research: State of MCP: October 2026

Frequently asked questions

Is there a hosted Chroma MCP endpoint?

Not in the README or MCP docs we opened. You run the server locally with uvx chroma-mcp. It can connect to Chroma Cloud as its storage backend.

Does my data survive a restart?

Not with the default ephemeral client, which keeps data only in memory. Use --client-type persistent with --data-dir, or the HTTP or cloud client, to keep it.

Which embedding models can I use?

The README lists default, cohere, openai, jina, voyageai and roboflow. External providers need an API key in an environment variable such as CHROMA_OPENAI_API_KEY.

Is the Chroma MCP server actively maintained?

As of November 2026, the latest GitHub release is v0.2.6 from August 2025, and PyPI shows the same version. The repo is not archived, but check recent activity before using it in production.

You Might Also like to Read

Get a FREE Website Audit

Enter your website and get a 20–30 page SEO, GEO, AEO & CRO audit within 24 hours. Completely free.

Don't have a site yet? Get in touch →

Ready to speak with a marketing expert?
Give us a call! (916) 866-7893

Websites Launched
1407
Websites Launched
Marketing Experts
29
Marketing Experts
Hours of Dedicated Work
204533
Hours of Dedicated Work
Irina S.Benjamin G.Jeanette M.David G.