// the find
ModernRelay/omnigraph
Lakehouse native graph engine with git-style workflows
Omnigraph is a graph database built on Lance columnar storage and S3-compatible object stores, designed for multi-agent AI workflows where many agents write to isolated branches and changes are merged back to main. It layers graph traversal, vector ANN search, and full-text search into a single query runtime, with Cedar policy enforcement on every write path. The target audience is teams building agent coordination infrastructure who need durable, versioned, auditable shared state.
The branch-per-agent model is the most interesting design decision here: agents write in isolation, changes are reviewable before merge, and the Lance format gives you time-travel queries over the graph history without extra infrastructure. The compiler crate being cleanly separated from the storage/runtime crate (omnigraph-engine) means the schema parser and typechecker have no Lance dependency, which is the right call for testability. Cedar policy enforcement happens at the engine level so the CLI, HTTP server, and embedded SDK all go through the same gate — you can't accidentally bypass it from one access path. The `cluster apply` idempotency and drift detection (plan/apply/validate) is a meaningful operational improvement over most embedded graph stores that have no concept of convergence.
Python SDK is listed as 'coming soon' with no ETA, which matters because most of the agent tooling ecosystem is Python-first — TypeScript-only client coverage is a real adoption barrier for the stated use case. The query language (.gq files with a custom PEG grammar) is another thing to learn and debug; it's not Cypher, SPARQL, or anything with existing tooling, and the linter is brand new. At 813 stars the community is small and the Slack is the only support channel — if you hit a Lance-level bug or a merge conflict edge case, you're reading source. There's no information on merge conflict semantics when two agent branches modify the same node: the README describes branching and merging but says nothing about what happens on true conflicts, which is exactly the failure mode that matters for concurrent agent writes.