# Omnigraph, source control for context graphs in Rust > Source: > Published: 2026-06-16 13:10:21+00:00 An agent reads the graph as context. The graph is how an agent understands the business. `$ curl -fsSL https://raw.githubusercontent.com/ModernRelay/omnigraph/main/scripts/install.sh | bash` Schema, context, policies, UI and infra are declarative files: versioned like code, enforced by the engine. Git-style branch, version, and merge for data. A set of linted, typed queries defines each lens on the graph, and branches keep concurrent batch writes safe. How it reads find_experts ranks people three ways at once (vector similarity, BM25 full-text, and graph structure) and fuses them with rrf(). The whole lens is a file you lint, version, and review. queries/team.gq ``` query find_experts($topic: String) { match { $p: Person } return { $p.name, $p.role } order { rrf(nearest($p.embedding, $topic), bm25($p.bio, $topic)) } limit 5 } ``` Many agents work the same graph at once, each on its own branch, merging into main only through review. Fig. A single hour on one graph: Hermes adds new data on hermes/ingest while Claude Code dedupes records on claude/dedupe. Claude Code and Hermes each fork main into their own branch. cheap fork`$ omnigraph branch create` Hermes adds new nodes and edges while Claude Code dedupes existing records. Isolated branches never collide. schema + policy enforced`$ omnigraph mutate` Each branch's diff is approved by a human or another agent before it lands. human or agent`$ omnigraph diff` Both land on main in review order, three-way and attributed to their agent. three-way · audited`$ omnigraph branch merge` 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. Vector, full-text, and graph traversal in one query, fused with rrf(). Writes are typed and checked, rejected at the branch before they land. Readers see a consistent snapshot, never a partial write. Row-level merges with typed conflicts, not last-write-wins. Every version stays queryable; pin a query to any snapshot. Destructive changes need explicit approval before apply runs. Data, catalog, and state live in one S3-compatible bucket. Every commit records its actor, resolved server-side. Denied and nonexistent are indistinguishable from outside. Your agents work with graph data the way engineers work with code: branched, typed, reviewed, merged, audited, and governed. An agent reads the graph as context. The graph is how an agent understands the business. Everything that defines it is a file. Schema, queries, policy, dashboards, deployment. Versioned like code. The discipline code already has. Everything you trust for code now governs knowledge. Agents move at a different speed. A team migrates a schema quarterly. A fleet wants to weekly. Plan. Apply. Review. Declarative files with a plan/apply cycle: the only form an agent can write and a human can review. An open platform, easy to integrate — available to agents, humans, and apps through different means. Typed client for omnigraph-server: instance-per-client, typed errors, async-iterator streaming export. @modernrelay/omnigraph ↗The graph as agent tools — schema, branches, queries, mutations and ingest, bridged into LLM hosts over stdio. @modernrelay/omnigraph-mcp ↗Declarative notebooks generated from the graph — lenses an agent can write and a human can read. Edit graph content as markdown files; changes flow back as reviewed commits. Omnigraph as the context layer for Hermes agents. Not options to weigh, but guides and blueprints to follow: docs to run locally, guides for your own servers, and clone-and-go blueprints for the cloud. A folder on your machine. graphs/*.omni, zero infra. Read the docs ↗Your server plus any S3-compatible store. Coming soonS3, IAM and KMS inside your own account. Clone the blueprint ↗One service booting from one bucket. Clone the blueprint ↗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. Track AI and ML market signals through the SPIKE frame: signals, patterns, insights, and know-how. Monitor trials, filings, competitors, and market movement in a graph built for competitive intelligence. Model decisions, traces, actors, and artifacts so an organization becomes queryable by agents. Connect deal flow, theses, founders, portfolio signals, and follow-up work. Track papers, claims, entities, and evidence across biotech and medical research. A personal operating graph for notes, people, commitments, routines, and recurring work. Built on a strong open-source foundation, aligned with the next-gen lakehouse graph. The engine: typed, fast, memory-safe systems code. Open columnar storage with versioning, so the graph remains readable files. The in-memory standard for columnar data interchange. Query execution for filtering, sorting, aggregation, and ranked retrieval. Policy enforcement in the engine, applied before writes land. On object storage you own — S3 or RustFS — in the open Lance format. No proprietary store, no lock-in. Agents branch the graph and propose writes in isolation; changes merge after review, like code — every change visible, reviewable and reversible. Yes. On-prem runs with no outbound calls, on your hardware and network; you hold every key. Through the TypeScript SDK against a running omnigraph-server, or via the MCP server that bridges Omnigraph into LLM hosts like Claude Desktop and Claude Code. Yes — MIT-licensed. The engine, CLI and server live in the ModernRelay/omnigraph workspace.