frontier-mcp

An MCP server that serves the markdown issue tracker under .scratch/ as a queryable graph — so agents stop re-parsing prose to learn what is open, blocked, or takeable.

Requirements

• Node 24 or later
• npx to run it; pnpm for development of this package

Install once (user scope)

Register FrontierMCP once in your user-level MCP config, pinned to a released version. Every repository you open then gets the server automatically — no per-repo mcp.json file.

In Cursor, edit ~/.cursor/mcp.json (create the file if it does not exist); other MCP clients take the same command and arguments in their own user-scope server list:

{
  "mcpServers": {
    "frontier": {
      "command": "npx",
      "args": ["-y", "[email protected]"]
    }
  }
}

Replace x.y.z with a published version — see the releases. The pin is the version you get; nothing bumps it for you.

Restart your editor after saving.

First use in a repository

1. Open the repository in your editor. FrontierMCP resolves the workspace from the session working directory — walking upward to the nearest .scratch/, or .git in either form — so opening the project is the only setup step. A git worktree is its own workspace: it carries .git as a file, and it is served instead of the repository it was made from.
2. Read the tracker configuration document once: fetch MCP resource frontier://tracker-doc, or read docs/agents/issue-tracker.md in a repo that vendors it.
3. Call list_efforts to see whether .scratch/ exists yet.
4. Call get_board on an Effort to see the Frontier, then get_tickets only for the ids you work.

A repository with no .scratch/ directory is not an error. Create the first Effort with create_tickets, edit_map, or spec and create: true.

Override the workspace

Optional, per call or per session:

• Pass root on any tool call to read another directory.
• Set FRONTIER_ROOT in the server environment for a non-standard layout.

The working directory is the one your client launched the server in, and it is fixed for the session. Moving to another worktree mid-session does not retarget it — use root, FRONTIER_ROOT, or restart the server.

Tools

Eight tools: list_efforts, get_board, get_tickets, create_tickets, update_ticket, edit_map, spec, migrate_effort. See the tracker configuration document for when to use each.

Development

pnpm install
pnpm test
pnpm run check
pnpm run build
node src/bin.ts

Releasing

Publishing is CI-only via release-it. There is no local publish script, and adding one would mean minting the long-lived npm token that the CI setup exists to avoid.

1. Merge the work you want to ship to master.
2. On GitHub: Actions → Release → Run workflow, pick auto / patch / minor / major. auto is the default and derives the bump from the conventional commits since the last release — feat: gives a minor, fix: a patch. Pick an explicit one to override it. If nothing since the last tag is releasable, auto exits green having shipped nothing, so confirm a new tag appeared.
3. The workflow runs checks + tests, bumps package.json, updates CHANGELOG.md, tags v*, creates a GitHub Release, and publishes to npm with pnpm.
4. Bump the pinned version in your user MCP config ([email protected]) when you want the new build — pins stay manual on purpose.

The workflow is dispatchable from any branch but refuses to run off master: release-it commits, tags and pushes as part of the run, so a release from a feature branch would rewrite that branch.

Local dry-run (no tag, no publish). It needs a clean working tree and an upstream branch, so commit first:

pnpm run release:dry

npm publishing uses Trusted Publishing — GitHub Actions OIDC, configured for this repo on npmjs.com, so there is no NPM_TOKEN to rotate.

Before changing .release-it.json or the workflow, read the release-it bullet in AGENTS.md Work Guidance. Several settings there look removable and are not — the absent registry-url, npm.skipChecks, npm.publishArgs, and the pnpm 10 pin each exist for a reason recorded in one place so it cannot drift.

License

MIT