Analytics cookies only (Google Analytics, PostHog). No ad networks, no cross-site tracking. See our privacy policy.
Self-host Chroma
updated Nov 2026deployed by usprices checked · Jul 2026Install verified on Ubuntu 26.04 · Nov 2026 · how we test
We earn commissions when you shop through the links below. Full disclosure →
An open-source vector database for AI applications: store embeddings with documents and metadata in collections, then query by similarity with metadata and document filters. It runs as a single server with Python, JavaScript/TypeScript and Rust clients and an HTTP API, persisting to local disk. Apache-2.0 licensed, with a hosted Chroma Cloud service from the same team.
Pick Chroma when you are building an AI application, such as retrieval-augmented generation, and want the smallest possible vector store to start with. You create collections, add embeddings with their documents and metadata, and query by similarity with metadata and document filters. The single-node server runs from one container with a Python, JavaScript/TypeScript or Rust client over HTTP, and persists to local disk. It is Apache-2.0 licensed, and the same team offers a hosted Chroma Cloud. Our snippet is the Docker deployment guide's `docker run` as a Compose file, with the image pinned to a release tag and a named volume on `/data`. Upstream's single-node performance page says to avoid deployments with less than 2 GB of RAM, which is where our 2048 MB floor comes from. The catalog's Qdrant is a Rust server with a REST API, gRPC and a dashboard, Weaviate adds modules for embedding providers, and Milvus is the multi-component option built for scale.
What you need
Any VPS with at least 2048 MB of RAM
A domain you control — most self-hosted setups need HTTPS in front of them
Chroma's Docker guide and single-node performance page flag these before you rely on it:
non-negotiableChroma's migration guide says it no longer provides built-in authentication, and the HTTP API on port 8000 is plain HTTP. Keep it on a private network or behind a reverse proxy that adds TLS and its own access control; anyone who can reach the port can read and write your collections.
non-negotiableMemory is the limit: the HNSW index must fit in RAM, and the performance page's rule of thumb is a maximum collection of about 0.245 million 1024-dimensional embeddings per GB of RAM, plus at least a gigabyte reserved for the system. A collection larger than memory slows sharply because the index layout does not swap well.
non-negotiableData lives in /data (SQLite plus the index files), which our chroma_data volume holds, so back that volume up. Disk guidance on the same page is storage at least as large as RAM plus several gigabytes.
non-negotiableThere is no web interface in the image; the guide shows access through the client libraries, and the heartbeat endpoint we use for readiness is /api/v2/heartbeat. We pin 1.5.9 (released 2026-05-05), although Docker Hub also carries newer dev tags that are not releases.
Chroma used 11 MB of RAM at idle on GCP e2-standard-2 · Ubuntu 26.04 LTS · Docker 29.9.3 (first boot, measured Nov 2026). We list a minimum of 2048 MB.
Is the Chroma install snippet verified?
The Chroma install snippet is marked verified on Ubuntu 26.04 (Nov 2026).