Sources: ModernRelay/omnigraph (MIT, about 2.2k stars), its README and the v0.15.0 release overview (released 9 Nov 2026), plus the MCP and policy guides in the repo. Surfaced by André Lindenberg on LinkedIn, 2026-11-13.
What Omnigraph is
An object-storage native graph database aimed at agents:
Typed ontology.
.pgschemas declare node and edge types, keys and constraints, enforced on every write.Git-style branches. A branch per agent or task, three-way merge, time travel. Agents propose reviewable changes instead of writing to main.
Open storage. Local disk or S3-compatible stores, data in the open Lance format. Vector, BM25 and graph retrieval in one engine.
Security as code. Cedar policy checked server-side on every mutation, per graph, per branch or server-wide, with bearer auth and audit.
Built-in MCP at
/mcp, plus a local bridge command.
The change in 0.15
In 0.14, MCP write tools were hidden unless the server ran with --mcp-allow-writes. But the same credential could make the same writes over HTTP. Hiding a tool from the MCP listing was presentation, not access control.
0.15 removes the flags. /mcp lists every graph operation (branch diffs, merge bases, publishing or dropping proposed commits, stored mutations), and each MCP tool runs its HTTP route's code under the same Cedar action. An agent gets the same answers and refusals over MCP as an HTTP client with the same credential. Only byte streams (blob payloads, exports) stay off MCP.
The upgrade trap
The release notes are explicit: if you serve MCP, agents gain every tool their policy allows, writes included, even if you never set a flag. A server started with the old flags refuses to start, and the old environment variables are silently ignored. To keep an agent read-only you must narrow its policy, and give it its own credential if it shares one with a person or service. An agent that signs in through OIDC acts as that person. A static token on a graph with no policy can still only read.
The local bridge keeps one client-side guard: without --write, it only admits tools annotated readOnlyHint: true. The server's policy still decides every call.
Other 0.15 changes worth knowing
Reads no longer wait for writes in the same process: the graph state is an immutable value behind an atomic pointer.
Diffs and merges cost what changed. After 1,024 one-row commits a fast-forward merge sends 54 storage requests instead of 4,168 (simulated store; vendor numbers).
Bulk CSV load of LDBC SNB SF30 in 87 to 110 s, down from 261 s (local disk, vendor numbers).
Single writer on S3 is now enforced by lease checks on every storage request; a server that loses the lease exits so the orchestrator restarts it.
Schema minimums like
@card(1..)are checked on every write, so a new node must arrive with its required edges.
The general lesson
Tool visibility is not authorization. Removing a tool from an agent's list lowers the chance it tries, but if the credential can reach the same operation another way, nothing is enforced.
Put the decision in one place, below every interface. One policy check shared by HTTP, CLI, SDK and MCP means one set of rules to audit.
Collapsing to policy can widen access on upgrade. Moving from "hidden by default" to "allowed by policy" is safer in the long run, but anyone who relied on the old hiding gets more access the moment they upgrade.
Review checklist
Does each agent have its own credential, not one shared with a human or service?
Is there an explicit policy for each agent, with writes scoped to its own branches?
Does any MCP-only restriction (hidden tools, client-side flags) also hold at the server?
Before upgrading, have you listed what each agent's credential can do under the new rules?
Caveats
Performance numbers are the project's own, mostly from a simulated store or local disk.
Cedar policy is only as good as what you write; the default for a graph with no policy and a static token is read-only, but other setups need deliberate rules.