HelixDB thinks your AI app needs one database, not five

What it does

HelixDB is an open-source database built from scratch in Rust that combines graph and vector search with key-value, document, and relational storage. It targets AI workloads—RAG pipelines, agents, and knowledge graphs—by letting developers query across data models in one place instead of wiring together separate vector, graph, and application databases. Queries are authored in Rust or TypeScript via SDKs and shipped as dynamic JSON to a local instance or an object-storage-backed cloud cluster.

The interesting bit

The project treats database sprawl for AI as the core problem. It bakes in embeddings, MCP tool discovery for agents, and graph traversals so an AI agent can walk relationships rather than generate human-readable queries. Local development uses a dynamic query API—POST a JSON AST to localhost:6969—so iterating doesn’t require a compile or deploy step.

Key highlights

• Multi-model engine: graph + vector are first-class, but KV, documents, and relational data are also supported.
• AI-native tooling: built-in text embedding, vector and keyword search, plus an MCP server that lets agents discover and traverse the graph.
• Rust and TypeScript SDKs share the same JSON query AST.
• Cloud tier runs on object storage with ACID transactions, a single writer, auto-scaling readers, and HA gateways.
• Licensed under the AGPL.

Caveats

• Local instances default to in-memory storage; stop the process and your data disappears unless you explicitly enable disk persistence.
• The cloud architecture uses a single writer, which caps write throughput regardless of how many readers scale out.
• The README touts “ultra-low latency” thanks to LMDB, but offers no published benchmarks to validate the claim.

Verdict

Grab it if you’re building RAG or agent systems and want one backend instead of a sprawling storage glue stack. Pass if you need a simple, single-purpose database or if AGPL is a non-starter.