Omnigraph MCP
A graph database with git-style branches: the server exposes it to agents over MCP at /mcp, and a skill teaches the agent schemas and queries
- 1,687 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The server listens on the network (`0.0.0.0:8080` in the example) and works with your storage, including S3 through `AWS_*` keys
- After the `--mcp-allow-writes` flag the agent changes the graph and creates, deletes and merges branches
- The install script runs through `curl | bash` from the main branch and puts binaries in `~/.local/bin`
Install
- Works in
Manual install
curl -fsSL https://raw.githubusercontent.com/ModernRelay/omnigraph/main/scripts/install.sh | bashInstalls `omnigraph` and `omnigraph-server` into `~/.local/bin` from published release binaries. The script comes from the main branch, so read it before running.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The server listens on the network (`0.0.0.0:8080` in the example) and works with your storage, including S3 through `AWS_*` keys
- After the `--mcp-allow-writes` flag the agent changes the graph and creates, deletes and merges branches
- The install script runs through `curl | bash` from the main branch and puts binaries in `~/.local/bin`
This is third-party code. Review the repository files before installing.
What it does
Omnigraph is a graph database with git-style branching. Data lives in the open Lance format on local disk or in S3-compatible storage. The `omnigraph-server` server exposes the database to agents over the Model Context Protocol at `/mcp`: no separate package is needed. It runs on Streamable HTTP, and each request carries its own token in the `Authorization: Bearer` header. Claude Code and Cursor connect with a URL and a token, and with OIDC trust configured the client signs in through OAuth. By default the agent gets read tools: `graphs`, `schema`, `queries`, `query` for stored queries, `snapshot`, `branches_list` and three calls for commit history. The `read` tool for ad-hoc GQ queries turns on with the `--mcp-allow-adhoc` flag, and the write tools `mutate`, `load`, `branches_create`, `branches_delete` and `branches_merge` with `--mcp-allow-writes`. Flags only reveal tools, and permissions come from the Cedar policy on the server: a call runs under the same policy action as a plain HTTP route. Besides MCP, the repository ships a skill for the agent. The command `omnigraph skill export ~/.claude/skills/omnigraph --force` writes a SKILL.md with guidance on `.pg` schemas, `.gq` queries, branches and policies, and its version matches the installed command line. The server serves the same reference material as `omnigraph://best-practices/` resources. Schemas and cluster deployment cannot be changed through MCP: the operator handles them with `omnigraph cluster plan` and `omnigraph cluster apply`.
Who it is for. For developers and teams building memory or a shared knowledge model for several agents who are ready to run their own server and storage.
Good fit when
- Several agents need a shared typed domain model with constraints rather than a pile of notes
- An agent should propose changes in a separate branch that you review and merge, as in git
- You need one query that combines graph traversal, vector search and full-text search
- Read and write rights must be set by a Cedar policy per graph and branch instead of trusting the agent
- Data must stay in your own storage: local disk, MinIO, S3, R2 or GCS
Not a fit when
- You need agent memory without a server and schemas: Omnigraph requires a `.pg` schema, a cluster and a running `omnigraph-server`
- You use a Mac with an Intel processor: the 0.13.0 release archives have none for it, leaving a build from Rust source
- You need to change a schema or deploy a cluster through the agent over MCP: only the operator can do that on the command line
- You want natural-language document queries without describing a schema: you must first describe node and edge types
Example request
Connect to my Omnigraph server, show the schema of the knowledge graph and find all sources that support the claim lower-latencyMCP
- Transport
- http
- Authentication
- API key
Environment variables
OMNIGRAPH_SERVER_BEARER_TOKENrequired, secret- A single static access token whose actions run as the default actor. A JSON object with several tokens can be used instead.
OMNIGRAPH_MCP_ALLOW_ADHOC- Set to 1 to turn on the read tool for ad-hoc GQ queries.
OMNIGRAPH_MCP_ALLOW_WRITES- Set to 1 to turn on the write and branch management tools.
OMNIGRAPH_MCP_ALLOWED_HOSTS- A comma-separated list of extra host names for the Host header check.
Limitations
- A request to `/mcp` must not exceed 64 KiB, a full tool result is up to 1 MiB, a read waits 30 seconds and up to 16 calls run at once. Large loads go through `POST /load/ndjson`.
- The server answers 403 when the host name is not allowed. Behind a load balancer without OIDC it must be added with the `--mcp-allowed-host` flag. With a plain static token the `graphs` tool also answers 403 until the cluster policy permits the `graph_list` action.
- A write that ran longer than 30 seconds or whose answer was lost returns `outcome_unknown`. Retry it with the `if_graph_commit` parameter, otherwise the change may apply twice.
- Ready release 0.13.0 archives exist for Linux x86_64 and arm64, macOS on arm64 and Windows x86_64. The skill describes version 0.14.0 and the HTTP contract 0.13: update the command line, the server and integrations together.
- MIT license. The repository front page still names the server package `@modernrelay/omnigraph-mcp`, while the MCP docs say the built-in `/mcp` replaced it.
How to disable. Remove the server from the MCP client settings and stop the `omnigraph-server` process. Remove the skill together with the `~/.claude/skills/omnigraph` folder, and the binaries from `~/.local/bin`. Graph data stays in your storage and is deleted separately.
FAQ
How do I connect Omnigraph to Claude Code or Cursor over MCP?
The server must know at least one credential: a static token from `OMNIGRAPH_SERVER_BEARER_TOKEN` or `OMNIGRAPH_SERVER_BEARER_TOKENS_JSON`, a signed one or OIDC. Without a token source `/mcp` is not mounted. In the client you enter a URL like `http://localhost:8080/mcp` and an `Authorization: Bearer` header; Claude Code and Cursor support this directly.
Can an agent write to the graph over MCP?
Only if the server runs with `--mcp-allow-writes`. Then `mutate`, `load` and branch operations become available, and each still passes the Cedar policy. It is safer to give the agent a separate branch and allow the `branches_merge` merge by hand.
What are stored queries and how does `query` differ from `read`?
A stored query is a `.gq` file in the cluster's `queries` folder that `cluster apply` registers. The `query` tool calls it by name with typed parameters. The `read` tool accepts arbitrary GQ text and is off by default.
How do I install the Omnigraph skill for an agent?
The command `omnigraph skill export ~/.claude/skills/omnigraph --force` writes a copy of the skill whose version matches your command line. For older versions without the `skill` command, `npx skills add ModernRelay/omnigraph@omnigraph` is suggested.
Where is data stored and can I use S3?
Data lives in Lance format on local disk or in S3-compatible storage: RustFS, MinIO, AWS S3, R2 or GCS. Azure Blob is still a preview. In a cluster the storage is set in `cluster.yaml` and keys are passed through `AWS_*` variables.
Discussion
No questions or reviews yet.
Related
MCP serversA CLI and MCP server that extracts structured knowledge from documents: graphs, hypergraphs, and fact search with sources
GraphJin MCP
GraphJin
GraphJin's MCP server gives an agent safe, schema-aware access to SQL and MongoDB instead of raw database credentials
An MCP server from AntV for generating charts and diagrams: 25+ visualization types, from lines and bars to maps, mind maps and network graphs
MCP server and CLI that turns paper text or data into academic figures, diagrams and plots
A Go MCP server and REST API for databases: PostgreSQL, MySQL, ClickHouse, Snowflake, BigQuery and more, with personal data masking and a cache
An MCP server with a toolbox for global briefs: markets, SEC filings, conflict, climate and news from open sources
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Omnigraph (this entry) | yes | API key | 1,687 | |
| Graphiti MCP | no | API key | 32,159 | agent memory as a time-aware knowledge graph on FalkorDB or Neo4j, with no branching or Cedar policies |
| MCP Knowledge Graph | no | not required | 886 | persistent memory in local .aim files, with no server, branches or typed schema |
| MCP Memory Service | no | not required | 2,036 | self-hosted memory with semantic search and auto-consolidation, with no graph branching |
README in short
The front page lays out the database's capabilities in a table: a typed ontology, search over graph, vectors and text in one query, branching, cluster as code, Cedar policies and object-storage persistence. Next come use cases, installation, a section on setup by an AI agent, three-step deployment, query and branch examples, clients and the build. The repository has user docs, change proposals, a governance guide, tests and a separate repository of ready example graphs.