index.md
Omnigraph
Object-storage native graph database with branching and typed ontology.
Your agents work with graph data the way engineers work with code: branched, typed, reviewed, merged, audited, and governed.
To set Omnigraph up, an agent can follow the onboarding guide: Onboarding Guide
In their words
I mostly use Omnigraph to build a company brain. It fit so well that I ended up contributing Azure Blob Storage support myself.
— Roey Zalta, Data & AI Solutions Engineer, Microsoft
Omnigraph is the right model for an agent-driven company brain. Agents work on branches across staging and production graphs, and nothing reaches main without review. It’s open source, so when I needed something, I sent a PR instead of waiting.
— Joan Reyero, VP of AI & Data, The Linux Foundation
Omnigraph is a foundation for
- Company brain. Org knowledge in one graph.
- Agent memory. Versioned memory per agent or task.
- Context graph. Decisions and tribal knowledge.
- Research graphs. Hypotheses and results, reviewed.
Key capabilities
Typed schema
All your agents share one model of the domain they work in, and every write is checked against it.
What to notice: A decision can't be saved unless it links to the meeting it was made in.
node Person {
handle: String @key
role: enum(engineer, pm, agent)
}
node Team {
name: String @key
}
node Meeting {
slug: String @key
title: String
held_on: Date
}
node Decision @instruction("Link every decision to the meeting it was made in.") {
slug: String @key
title: String
body: String @index
embedding: Vector(1536)? @embed("body") @index
}
edge MemberOf: Person -> Team @card(1..1)
edge Attended: Person -> Meeting
edge DecidedIn: Decision -> Meeting @card(1..)
Branches
Agents work on their own branch, so nothing they write reaches the shared graph until you've reviewed exactly what changed and merged it.
What to notice: While scout works, main doesn't change. Its work only lands once you've seen the diff and merged.
# scout works on its own branch
$ omnigraph branch create agent/scout
$ omnigraph mutate --branch agent/scout --as scout -e '
query decide() {
insert Decision { slug: "dead-letter", title: "Dead-letter failed webhooks",
body: "Replay them by hand after the last retry." }
insert DecidedIn { from: "dead-letter", to: "webhook-review" }
}'
changed agent/scout via decide: 1 nodes, 1 edges
# main doesn't see it yet
$ omnigraph query --branch main -e '
query check() { match { $d: Decision { slug: "dead-letter" } } return { $d.title } }'
0 rows from branch main via check
# when the work is done: review, then merge
$ omnigraph branch diff agent/scout
insert node Decision dead-letter
insert edge DecidedIn dead-letter -> webhook-review
$ omnigraph branch merge agent/scout --into main
merged agent/scout into main: fast_forward
Hybrid search
Agents get the right context in one query: follow relationships, filter on structured fields, and rank by meaning and keywords.
What to notice: Vector search alone would also return Tom's retry decision. Walking the graph first keeps only decisions from the platform team's meetings.
query team_decisions($team: String, $q: String) {
match {
$t: Team { name: $team }
$p memberOf $t
$p attended $m
$d decidedIn $m
}
return {
$d.title, $m.title,
rrf(nearest($d.embedding, $q), bm25($d.body, $q)) as score
}
order { rrf(nearest($d.embedding, $q), bm25($d.body, $q)) }
limit 5
}
$ omnigraph query team_decisions --query search.gq \
--params '{"team":"platform","q":"how many times do we retry webhooks?"}'
2 rows from branch main via team_decisions
d.title | m.title | score
----------------------------------------+----------------------------+-------
Retry webhooks with exponential backoff | Webhook reliability review | 0.0328
Add a retry budget per customer | Queue outage postmortem | 0.0323
Declarative config
The whole deployment is a folder in Git: the graphs, their schemas, the queries agents can call, and who may do what. You change it like code, and a plan shows what will happen before anything does.
What to notice: The plan lists every change before anything is written, down to the schema migration, so a change to what agents can do gets reviewed like a code change.
version: 1
graphs:
brain:
schema: schema.pg
queries: queries/
policies:
brain-access:
file: graph.policy.yaml
applies_to: [brain]
# schema.pg gives decisions a status; queries/open.gq is new
$ omnigraph cluster plan --target ../company-data --as alice
cluster plan: 3 change(s)
Update graph.brain
Create query.brain.open_decisions
Update schema.brain
{"kind":"add_property","type_name":"Decision","property_name":"status",…}
$ omnigraph cluster apply --target ../company-data --as alice --json \
| jq -c '{status, converged: .result.converged}'
{"status":"complete","converged":true}
Object storage
The whole graph lives in object storage as open Lance tables: the data, every branch and commit, and the config. Servers only compute, so there's no database disk to size or grow, and the graph gets the durability, scale and cost of object storage. The files are open, so DuckDB or any other Lance reader can query them too.
What to notice: No server keeps anything of its own. Add readers as traffic grows, and a new writer picks up every branch and commit where the last one stopped.
Where the graph lives:
agents and apps
|
v
+- Omnigraph servers, stateless ---+ your tools
| +--------+ +--------+ +--------+ | +--------+
| | reader | | reader | | writer | | | DuckDB |
| +--------+ +--------+ +--------+ | +--------+
+----------------------------------+
| | | |
v v v v
+--------------------------------------------------+
| s3://your-bucket/ |
| __cluster/ queries, policies |
| graphs/company.omni/ |
| __manifest/ branches |
| __history/ commits, schemas |
| nodes/ |
| 0000...000a/ Person, a Lance table |
| data/*.lance rows, as columns |
| _indices/ vector, FTS, scalar |
| _versions/ every version |
| 0000...0009/ Meeting |
| edges/ |
| 0000...000b/ Attended |
+--------------------------------------------------+
The Knowledge IDE
A blazing fast, GPU-rendered desktop app to visualize your ontology and graph, inspect knowledge, and review changes before they merge.
- Spot changes. Open an agent's branch and see what it added, changed and removed, in place.
- Inspect a record. See any entity as a page: its properties and everything linked to it.
- Explore links. Follow relationships out on a canvas or down a tree.
- Query safely. Write queries against the schema. A typo shows up before anything runs.
- Approve a merge. Go through a branch change by change, then merge it into main.
Git-style workflows
Agents safely collaborate and write to the same graph, each on its own branch, merging into main only through review.
No agent touches main directly. The schema rejects malformed writes at the branch, the policy decides who can merge, and the commit graph records every change and its author. Two agents, one graph, no collisions.
Deploy anywhere
- Embedded. A folder on your machine. No server, no infra.
- BYOC. AWS, Azure or GCP, in your own cloud account.
- On-prem. Your own servers and any S3-compatible store.
- Omnigraph Cloud. Run by us: a dedicated cluster with SSO, policy and audit.
Templates
Cookbooks are runnable Omnigraph projects: schema, queries, seed data, and agent skills in one repo. Clone one, apply it, and inspect a real graph before you design your own.
- Industry intel. Track AI and ML market signals through the SPIKE frame: signals, patterns, insights, and know-how.
- Pharma intel. Monitor trials, filings, competitors, and market movement in a graph built for competitive intelligence.
- Company brain. Claims, decisions, and action items with the exact source text they rest on, shared by a company's people and agents.
- VC OS. Connect deal flow, theses, founders, portfolio signals, and follow-up work.
- Beads. A beads-style work queue for coding agents, driven from the ob client.
- Second brain. A personal operating graph for notes, people, commitments, routines, and recurring work.
Foundation
Built on a strong open-source foundation, aligned with the next-gen lakehouse graph.
At a glance
| Engine | Rust |
| Version | 0.15.0 |
| Storage | S3-compatible · Azure Blob (preview) · local disk |
| Format | Lance · open columnar |
| Retrieval | graph · vector · full-text · hybrid (RRF) |
| Branching | zero-copy · git-style · three-way merge · time travel |
| Policy | Cedar · enforced in-engine |
| Config | Terraform-style · cluster + operator surfaces |
| Clients | CLI · TypeScript SDK · Python SDK · MCP |
| License | MIT · open source |
| Deploy | embedded · BYOC (AWS, Azure, GCP) · on-prem |
FAQ
What is a context graph?
A graph of what your agents need to know: people, projects, decisions, documents, and how they connect, kept current as agents and people work. Omnigraph is the database underneath one: typed, versioned, and searchable by graph traversal, vector and full-text in a single query.
Is Omnigraph an open source graph database?
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.
Graph database vs vector database: which do AI agents need?
Usually both. A vector database finds text similar to a question; a graph database follows how people, projects and decisions connect. Omnigraph does both in one query: vector, keyword and hybrid ranking over the same typed nodes you traverse, so there are no two stores to keep in sync.
Can I use Omnigraph for GraphRAG?
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 the related entities and sources, so the context you give the model comes with its connections.
What database should I use for AI agent memory?
One your agents can write to safely and you can audit. In Omnigraph each agent writes on its own branch, the schema is enforced on every write, Cedar policies run inside the engine, and changes reach main only through a merge you can diff first. Every version stays readable, so you can see what an agent knew and when.
Is Omnigraph a Neo4j alternative?
For agent workloads, yes. Neo4j is a mature graph database. Omnigraph is MIT-licensed, keeps the graph as open files in storage you own, and is built around what agents need: git-style branches and merges, a typed schema checked on every write, built-in vector and full-text search, and in-engine policy.
Does Omnigraph work with MCP?
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 it, write. omnigraph mcp connects a local agent to a served graph.
Where does my data live, and what does it cost?
In storage you own (S3-compatible, Azure Blob or local disk), as open Lance files that stay readable without us. The engine is free and open source: you pay only for your own storage and compute.
Get started
Install the omnigraph CLI on macOS or Linux:
curl -fsSL https://raw.githubusercontent.com/ModernRelay/omnigraph/main/scripts/install.sh | bash
or with Homebrew:
brew tap ModernRelay/tap
brew install ModernRelay/tap/omnigraph