Chroma

Chroma is the developer-first open-source vector database. Its API is intentionally minimal — collection.add(documents=...) handles embedding, indexing, and storage in one call — which makes it the fastest path from idea to a working RAG prototype.

For agents, Chroma's appeal is the same code path running on a laptop, a single server, or the managed Chroma Cloud. The same Python or JavaScript client talks to all three. Built-in embedding functions cover OpenAI, Cohere, Voyage, Hugging Face, and local SentenceTransformers, so you don't wire up a separate embeddings service for prototypes.

Chroma Cloud introduces serverless billing with separated storage and compute, plus a first-party MCP server that exposes collection and document operations as agent tools. Tenant and database scoping provides multi-tenant isolation without running multiple clusters.

The tradeoffs are scale and tunability. Chroma uses HNSW only — no IVF, PQ, or DiskANN — and is younger at billion-vector scale than Milvus or Pinecone. For prototypes, internal tools, and small-to-medium production RAG, those limits rarely bite; for very large or latency-critical workloads, look at Qdrant or Milvus.

Agent-specific features:

• One-line ingest with built-in embedding functions across major providers
• Tenant and database scoping for multi-tenant agent apps
• Official MCP server exposing collections and documents as agent tools
• Chroma Cloud serverless with usage-based billing
• Identical local and cloud APIs — no rewrite when graduating from prototype