Open source graph database for AI agents and context graphs
Omnigraph is an open source (MIT-licensed) graph database aimed at AI agents and "context graphs" — company brains, agent memory, and decision records. Its engine is written in Rust and stores the graph as open Lance columnar files in S3-compatible storage, Azure Blob (preview), or a local directory, so the data lives in infrastructure you own. Its distinguishing model is Git-style branching: agents each work on their own branch, writes are validated against a typed schema, Cedar policies run in-engine, and changes reach main only through a reviewed three-way merge. Retrieval combines graph traversal, vector, full-text, and hybrid (RRF) search in a single query.
Define nodes, edges, keys, enums, and cardinality constraints in a schema file; malformed writes are rejected at the branch. A decision, for example, can't be saved unless it links to the meeting it was made in.
Agents fork main into their own branches with zero-copy forks, write in isolation without collisions, and merge through review. The commit graph records every change and its author, with time travel to prior versions.
Vector, keyword, and hybrid (RRF) ranking over the same typed nodes you traverse — no separate vector store to keep in sync. Supports @embed directives on fields (e.g. Vector(1536)).
Terraform-style configuration with cluster and operator surfaces; Cedar policies decide who can merge and are enforced inside the engine.
The graph is stored as open Lance files in S3-compatible storage, Azure Blob (preview), or a local folder — readable without Omnigraph. Deploys embedded (no server), BYOC on AWS/Azure/GCP, on-prem, or via Omnigraph Cloud.
A GPU-rendered app for macOS, Linux, and Windows to visualize the ontology and graph, spot what an agent's branch changed in place, inspect entities as pages, query with schema-checked typing, and approve merges change by change.
CLI plus TypeScript and Python SDKs. omnigraph-server serves MCP at /mcp so Claude, ChatGPT, Cursor, or any MCP client can read the schema, run queries, and (where Cedar policy allows) write. The site shows integration with agents like Claude Code and Hermes working on parallel branches.
Six starter projects combining schema, queries, seed data, and agent skills — including company brain, industry and pharma intel, VC OS, a beads-style work queue, and a personal second brain.
Declare nodes, edges, keys, and constraints in a schema file (schema.pg); every write across all agents is checked against this shared domain model.
Each agent (human or LLM) forks main into its own branch with a zero-copy fork, then adds, changes, or dedupes nodes and edges in isolation. The schema and policy are enforced on every mutation.
Each branch's diff is approved by a human or another agent via omnigraph diff, then merged into main three-way, attributed to its author, and audited in the commit graph. No agent touches main directly.
Run a single query that traverses typed nodes and ranks results by vector, keyword, or hybrid (RRF) similarity — the basis for GraphRAG and agent memory use cases.
The engine is free and open source (MIT): you pay only for your own storage and compute. Data lives in storage you own (S3-compatible, Azure Blob, or local disk) as open Lance files. Omnigraph Cloud (a dedicated managed cluster with SSO, policy, and audit) is in private preview; no pricing is listed.
An open source, object-storage-native graph database for AI agents and context graphs. It combines a typed ontology, Git-style branching with three-way merges, hybrid retrieval (graph, vector, full-text, RRF), and in-engine Cedar policy, storing data as open Lance files in storage you own.
Yes. The engine, CLI, and server are MIT-licensed on GitHub. The graph is stored as open Lance files in S3-compatible storage, Azure Blob (in preview), or a local directory, so it runs anywhere — embedded on a laptop with no server, or in your own cloud account.
The engine is free and open source; you pay only for your own storage and compute. Omnigraph Cloud (dedicated managed cluster with SSO, policy, and audit) is currently in private preview.
In storage you own — S3-compatible storage, Azure Blob (preview), or a local disk — as open Lance files that remain readable without Omnigraph. Deployment options are embedded (a folder on your machine, no server), BYOC (AWS, Azure, or GCP in your own account), on-prem, or Omnigraph Cloud.
Yes. Store documents, chunks, and the entities they mention as typed nodes and edges with embeddings on the nodes. One query ranks chunks by vector, keyword, or hybrid similarity and traverses to related entities and sources, so context comes with its connections.
Each agent works on its own branch. The schema rejects malformed writes at the branch, Cedar policies decide who can merge, and changes reach main only through a diffable, attributed three-way merge. Every version stays readable, so you can see what an agent knew and when.
For agent workloads, yes. It is MIT-licensed, keeps the graph as open files in storage you own, and is built around agent needs: git-style branches and merges, schema checked on every write, built-in vector and full-text search, and in-engine policy.
Yes. omnigraph-server serves MCP at /mcp, so Claude, ChatGPT, Cursor, or any MCP client can read the schema, run queries, and — where your Cedar policy allows — write. omnigraph mcp connects a local agent to a served graph.