Genie MCP server

A standalone server that connects any MCP-aware client, Claude Desktop, Cursor, Continue, Zed, to your Genie org. It reflects Genie's tool registry, so the same read tools the chat uses show up natively in your editor or desktop agent.

What it is

@genie/mcp-server is an Apache-2.0 npm package that speaks the Model Context Protocol. It runs as a child process spawned by your MCP client and talks to Genie over HTTP using an org API key. It does not need the Genie cloud to run locally, just an account and a key.

It is not a fixed, hand-written tool list. At startup it calls GET /api/v1/tools to fetch the org-plane tools your cloud exposes, registers each one as an MCP tool, and forwards calls to POST /api/v1/tools/{name}. Add a tool to the registry and it appears in your MCP client on the next restart, no client config change, no drift.

Install & configure

Most clients want a command string rather than an installed binary, so npx is the canonical invocation. Point it at your Genie API base and a Bearer API key (see Authentication to mint one):

json
{
  "mcpServers": {
    "genie": {
      "command": "npx",
      "args": ["-y", "@genie/mcp-server"],
      "env": {
        "GENIE_API_BASE": "https://dev.genie.tech",
        "GENIE_API_KEY": "sk-genie-..."
      }
    }
  }
}

The server refuses to start without both GENIE_API_BASE and GENIE_API_KEY. The key is org-scoped: every tool call is attributed to the key's creator and runs against that org's data only.

What you can do over MCP

The MCP surface is org-plane reads. Once connected, your client can call the reflected registry tools and the built-in context tools:

Read your verification & project data

• list_proof_packets / get_proof_packet - browse and read proof packets.
• list_proof_artifacts / read_proof_artifact - pull the evidence inside a packet.
• get_review - read a completed review by its job id.
• get_rollup - the latest rollup digest for your org.
• list_projects, list_wishes, check_wish_done, list_recent_jobs, get_usage_summary.
• exec_cortex_search and the other Cortex read tools - query your project's facts and hotspots.
• search_genie_docs, read_doc, and (when configured) web_search.

Shared context (cross-machine clipboard)

• context_save - persist a snippet under an optional stable name.
• context_get - fetch by ctx_… id or name.
• context_list_recent - list recent entries.
• Resources context://recent and context://named/<name> are auto-discoverable. See Genie Context.

What it does not do (yet)

Specifically, you cannot start a proof packet or a review from an MCP client today:

• Triggering a proof packet is a control-plane action (trigger_proof_gen is marked destructive: it spends credits and queues GPU work). Control-plane tools stay behind the in-app confirmation card in Genie Chat, which needs a human in the loop, so they are filtered out of the MCP surface. It is also keyed to a wish or project, not a PR or commit.
• Triggering a review has no tool at all, over MCP or in chat. Reviews are started by the Genie GitHub App on a PR; the registry only exposes get_reviewto read the result. So "run a review on this PR/commit from my editor" is not yet supported.
• Admin actions like mint_api_key are admin-only and also excluded.

In short: MCP today lets an agent read your proofs, reviews, rollups, and facts; it cannot kick them off. Read-then-act flows still route through chat or the GitHub App.