The MCP server for GitHub org context

The Model Context Protocol (MCP) is how agents pick up tools. The EOS MCP server turns your whole GitHub org into a set of native tools any MCP-compatible agent — Claude, Cursor, or your own — can call, with zero glue code. Here’s how it works and how to wire it up in a minute.

Install it

There’s nothing to host. The server runs over stdio via npx; it just needs your API key. In any MCP client, add:

{
  "mcpServers": {
    "eos": {
      "command": "npx",
      "args": ["-y", "@eos-ai/mcp-server"],
      "env": { "EOS_API_KEY": "eos_your_key" }
    }
  }
}

That’s the same block for Claude Desktop, Cursor, and most MCP clients — drop it in the client’s MCP config, restart, and the tools appear. Grab a key from the console first (Settings → API).

The seven tools

Each tool maps to a context surface and returns cited, LLM-ready JSON:

• ›search_context — natural-language retrieval over the org, with citations to files, PRs, and people.
• ›get_ownership — ranked owners of a file or module, plus a knowledge-concentration (bus-factor) signal.
• ›get_graph — modules and inter-file dependencies across the org.
• ›get_coupling — files that change together (logical coupling) — blast radius.
• ›get_contributors — per-person expertise and review history.
• ›suggest_reviewers — scored reviewer suggestions for a pull request.
• ›get_activity — where change is concentrated right now — hotspots and churn.

What it looks like in use

Once the server is connected, the agent calls tools on its own. Ask it something org-shaped and it reaches for the right one:

You:   Who should review my change to services/auth in acme/app?
Agent: → calls suggest_reviewers / get_ownership
       Sam Lee is the primary owner (84% of recent changes,
       47 reviews in 90d). Backup: Ren Park. Sam has no open
       PRs right now — good pick. [cites blame + review graph]

No prompt engineering to teach it your org, no scraping the GitHub API in a tool wrapper — the context comes back pre-computed and cited, so the agent can ground its answer instead of guessing.

Prefer code?

The MCP server is a thin wrapper over the same REST API. If you’re building rather than chatting, the TypeScript SDK exposes every surface directly:

import { EosClient } from "@eos-ai/sdk";

const eos = new EosClient({ apiKey: process.env.EOS_API_KEY });
const { data } = await eos.context.reviewers(482, { repo: "acme/app" });

See the API reference for every endpoint, or read what a GitHub Org Context API is for the bigger picture.

Try it

Free tier: 5,000 context requests a month, no credit card — the full surface and the MCP server included. Add the config above, restart your agent, and ask it who owns something.