pm-groundwork-mcp

The MCP server for PM Groundwork, a starter kit that gives project and product managers an AI assistant that remembers their project between sessions.

This package is for Cursor, Codex CLI, and Gemini CLI. If you use Claude Code, Claude Desktop, or Claude Cowork, install the plugin instead — you don't need this package.

Claude Code:

/plugin marketplace add zoiestar/pm-groundwork
/plugin install pm-groundwork

Claude Desktop or Cowork: Customize → Plugins → Personal plugins → + → Add marketplace → Add from a repository → https://github.com/zoiestar/pm-groundwork. In Cowork, open the Cowork tab first.

Quick start

In your project folder:

npx pm-groundwork-mcp init

This detects which AI tools you have installed and writes their configuration. Restart your tool afterward, then run the pm-setup prompt.

Manual configuration

Cursor — .cursor/mcp.json

{ "mcpServers": { "pm-groundwork": { "command": "npx", "args": ["-y", "pm-groundwork-mcp"] } } }

Gemini CLI — .gemini/settings.json

{ "mcpServers": { "pm-groundwork": { "command": "npx", "args": ["-y", "pm-groundwork-mcp"] } } }

Codex CLI — ~/.codex/config.toml

[mcp_servers.pm-groundwork]
command = "npx"
args = ["-y", "pm-groundwork-mcp"]

What you get

21 prompts, generated from the same source as the Claude Code skills so behavior matches across tools:

| Prompt | What it does | |---|---| | pm-setup | Builds the workspace. Shows what it will create and waits for your yes. | | pm-project-start | Pick documentation, prototype, or full delivery. Builds the plan and the document set. | | pm-start-session | Briefs you on status, blockers, decisions due, and what changed in git. | | pm-checkpoint | Mid-session save. No commit. | | pm-end-session | Daily log, plan update, memory, secret scan, commit, push. | | pm-draft | Choose from 14 document types. | | pm-draft-prd, pm-draft-project-charter, and 13 more | Skip the chooser and draft one type directly. |

8 tools for reading and writing the workspace: pm_read_workspace, pm_write_workspace_file, pm_update_workspace_file, pm_log_decision, pm_write_daily_log, pm_scan_workspace, pm_scan_project_files, pm_check_decisions_due.

5 resources: pm://workspace/context, memory, decisions, user, and a computed status.

Notes

Workspace layout is detected automatically. Projects set up by v1, v2, or v3 of PM Groundwork all work; files are read from wherever they actually are and written back to the same place.

Questions adapt to your client. Tools without a structured question interface get numbered lists and an instruction to wait for your answer.

Where your data goes: nowhere. Everything is plain markdown files in your project folder. There is no service behind this.

Security

This server speaks stdio only — it is launched by your AI tool as a local subprocess and never opens a network port.

npm audit reports advisories in the Model Context Protocol SDK's HTTP server dependencies (hono, express-rate-limit, body-parser, qs). Those are transitive, they belong to the SDK's HTTP transport, and this package does not use that transport — the affected code never runs here. They are reported because npm audits the whole dependency tree, not the code paths you actually execute.

The server reads and writes only inside your workspace directory. Paths that would escape it are rejected.

Development

npm install
npm test     # builds, then runs 20 protocol conformance checks

Prompts are generated from the skills/ directory at the repository root. Never edit src/generated/ — npm run check:parity fails the build when generated prompts drift from their source.

Full documentation: github.com/zoiestar/pm-groundwork

License

MIT