org-context-mcp

An MCP server that gives AI agents your organisation's context — vision, products, teams, working methods — straight from a git repository you own, following the Org Context Spec. Agents read the context before they work, and propose changes as pull requests that a human reviews: the model recommends, a human decides.

Endpoint: POST /mcp (streamable HTTP). Nothing is stored server-side — your repo and token travel in the headers of each request.

1 · Create your context repository

Any GitHub repository (private is fine) holding markdown files in this structure at its root:

README.md              ← reading order and writing rules; agents start here
organisation/          ← vision, goals
products/<area>/product.md   ← per product: vision, effect goals, audience
products/<area>/actors/      ← actor files with assumptions
teams/                 ← per team: what it builds, tracker mapping
method/                ← how you write epics, stories, hypotheses
system/                ← pointers to the systems that own the details

Start small — a README and one product file is enough to be useful. Or start from nothing: connect an empty repository and ask your AI to set it up — it gets the spec's starter templates from the server, interviews you about what you know so far, and seeds the repo through the same review loop that later maintains it.

2 · Connect

Install the GitHub App

Install it on your context repository only. You come back here with a ready-to-run command per repository — and nothing is stored on the server: the key in the command is derived, not saved. Revoke it any time by uninstalling the App.

Alternative: a fine-grained access token

On GitHub: Settings → Developer settings → Fine-grained tokens → Generate new token.

  1. Repository access: Only select repositories → pick your context repo alone. That way the token can never touch anything else.
  2. Permissions → Repository permissions:
    • Contents: Read and write — to read the context and commit changes.
    • Pull requests: Read and write — every change becomes a pull request for human review.

Both permissions are needed. With Contents alone, a proposal gets committed to a branch but the pull request fails with 403.

Then fill in the form — the command is generated locally in this page, nothing is sent anywhere:

Then run this in your terminal to connect Claude Code:

Other MCP clients (Claude Desktop, VS Code, Cursor, …): configure an HTTP server with the same URL and the three headers shown above.

The full specification, starter template, and a worked example live in the Org Context Spec repository.