Omnigraph

Omnigraph is an open-source, MIT-licensed graph engine from Modern Relay, built around the Lance format and object storage. It gives agents a typed, versioned graph they can read as context and write to through reviewable branches. It ships a CLI, a server, a TypeScript SDK and an MCP server, and a hosted Omnigraph Cloud is listed as a private-preview waitlist.

What It Is

Omnigraph is a lakehouse-native graph database designed for context assembly and as an operational state and coordination layer for agents. Schemas (.pg files) declare node and edge types, keys and constraints, and the engine enforces them on every write. Data, catalog and state live in a local folder or any S3-compatible bucket in the open Lance format.

Git-style workflows for agents

Each agent or task can fork its own branch from main, write typed changes in isolation, and have the diff reviewed by a human or another agent before a three-way merge. Commits are atomic across the whole graph, every version stays queryable through time travel, and each commit records its actor, resolved server-side. Destructive changes can require explicit approval before apply runs.

Retrieval and declarative configuration

Queries are stored, linted .gq files. A single query can combine vector similarity, BM25 full-text and graph traversal, fused with Reciprocal Rank Fusion via rrf(). A cluster.yaml declares graphs, schemas, stored queries, embedding providers and policies, and the cluster plan and cluster apply commands preview and converge changes in a Terraform-style loop. Cedar policies are enforced in-engine on writes.

Deployment and setup path

Install with a curl script or Homebrew, which provides the omnigraph CLI and omnigraph-server. Run it locally, on your own servers with an S3-compatible store, or follow AWS and Railway blueprints in the cookbooks repository. The cookbooks provide runnable example graphs such as industry intel, pharma intel, context graph, VC OS, biomed research and second brain. The engine is written in Rust on Arrow, DataFusion, Lance and Cedar.