Omnigraph – A Graph Database Built for AI Agents

What it is

Omnigraph is an object‑storage‑native graph database that stores its data in the open‑source Lance columnar format. It is designed for the kind of long‑running, multi‑agent workloads that modern LLM‑powered systems need: a shared, typed world model that can be queried with graph traversal, vector similarity, and full‑text search all in a single request.

Core ideas

Idea	Why it matters for AI/LLM workloads
Typed ontology	Schemas (.pg files) define node/edge types, keys and constraints. Agents can rely on a single, enforced model of the domain, reducing “hallucinated” or malformed data.
Multimodal retrieval	A single query can combine graph hops, ANN vector search, BM25 text search and rank‑fusion, letting agents pull the most relevant context in one go.
Branching & versioning	Each agent or task can work on its own branch, submit a review, and merge back. The whole graph is time‑travelable, which is useful for audit, reproducibility, and “agentic memory”.
Infrastructure as code	A cluster.yaml declares graphs, schemas, stored queries, embedding providers and security policies. Commands like cluster plan / cluster apply behave like Terraform, giving a predictable, repeatable deployment flow.
Security as code	Cedar policies are stored alongside the data and enforced on every mutation, with bearer‑token authentication and audit tracking.
Object‑storage native	Data lives on any S3‑compatible store (AWS S3, MinIO, Cloudflare R2, GCS, Azure Blob preview) or on‑premise RustFS. Because Lance is a columnar, versioned format, storage costs stay low and the system can scale horizontally.

Typical use‑cases

• Company brain – a single knowledge graph that all internal agents can query.
• Agentic memory – durable, branch‑based memory for each autonomous agent.
• Context graph – capture decision traces and tribal knowledge for retrieval.
• Dev graph – model issues, dependencies and code artefacts that coding agents read/write.
• R&D / ML data layer – versioned experiment data that can be fed directly into training pipelines.

How you interact with it

Tool	What it does
omnigraph (CLI)	Create/validate clusters, run stored queries, mutate data, manage branches.
omnigraph-server	HTTP server exposing a REST/OpenAPI interface; each graph is reachable at /graphs/{id}/….
TypeScript SDK (@modernrelay/omnigraph)	Typed client library for Node/TS applications.
MCP server (@modernrelay/omnigraph-mcp)	Bridge that lets LLM hosts (Claude, Codex, etc.) call Omnigraph via a simple protocol.
Python SDK (planned)	Future typed access from Python.

Quick start (from the README)

# install the CLI and server
curl -fsSL https://raw.githubusercontent.com/ModernRelay/omnigraph/main/scripts/install.sh | bash
# or via Homebrew
brew tap ModernRelay/tap && brew install ModernRelay/tap/omnigraph

# declare a cluster directory
mkdir -p company-brain && cd company-brain
cat > cluster.yaml <<'EOF'
version: 1
metadata:
  name: company-brain
storage: s3://my-bucket/company-brain
graphs:
  knowledge:
    schema: people.pg
    queries: queries/
policies:
  base:
    file: base.policy.yaml
    applies_to: [knowledge]
EOF

# validate, plan, apply
omnigraph cluster validate
omnigraph cluster plan
omnigraph cluster apply

# start the server
omnigraph-server --cluster . --bind 0.0.0.0:8080

You can then run queries or mutations, e.g.:

omnigraph query search_docs --params '{"q":"AI safety"}'
omnigraph mutate add_person --params '{"name":"Mina"}'

Branching works similarly:

omnigraph branch create --from main agent/ingest-42
omnigraph branch merge agent/ingest-42 --into main

Security model

• All writes pass through a Cedar policy bundle defined in the cluster.
• Bearer tokens are hashed at startup; the server resolves the actor identity, preventing clients from spoofing it.
• Policies can be scoped per‑graph, per‑branch, or server‑wide, giving fine‑grained governance.

Who might use it?

• Teams building autonomous AI assistants that need a shared, versioned world model.
• Companies wanting a central knowledge graph that can be queried with vector search for LLM‑augmented retrieval.
• Researchers needing a low‑cost, versioned data lake for experiment metadata that can be accessed via graph or vector queries.

Where to go next?

• Explore the cookbooks repository for ready‑made domain graphs (company brain, VC ops, pharma intel).
• Read the detailed cluster guide and deployment guide for production‑grade setups (Kubernetes, AWS, Azure, etc.).
• Join the Omnigraph Slack community for support and roadmap discussions.

---

All information above is taken directly from the project's README; no additional features have been inferred.