parapetai-mcp¶

parapetai-mcp is an MCP server for Claude Code (and any other MCP client): device-code login to a Parapet control plane, provisioning agents, pulling the exact quickstart config a codebase needs, and six packaged skills that scaffold, retrofit, or audit a governed agent project for you.

It ships as its own PyPI package, separate from parapetai-agent — you don't need it to use the SDK, and using it never requires writing Cedar-aware code by hand if you'd rather have an agent do it.

Install¶

It's a CLI tool (a [project.scripts] entry point), not a library anything imports — and most Python installs on macOS/Linux now ship as PEP 668 "externally managed", which makes a bare pip install parapetai-mcp fail. Use pipx:

pipx install parapetai-mcp

No pipx? brew install pipx on macOS, or python3 -m pip install --user pipx && pipx ensurepath elsewhere. If uv is already on the machine, uv tool install parapetai-mcp is an equivalent one-liner.

Requires Python 3.12+.

Set up a project¶

parapetai-mcp init                 # installs all packaged parapet-* skills into .claude/skills/
claude mcp add parapet -e PARAPETAI_CONTROL_PLANE_URL=https://app.parapet.run -- parapetai-mcp serve

init [project_dir] (default .) copies every packaged skill — parapet-maf, parapet-adk, parapet-langgraph, parapet-quickdemo, parapet-install-prereqs, parapet-audit, parapet-audit-fix — into <project_dir>/.claude/skills/parapet-<name>/. It never touches anything outside .claude/skills/; it does not instrument your project's code itself. That's what the skills themselves do once invoked.
claude mcp add ... registers parapetai-mcp serve as an MCP server with Claude Code, pointed at a control plane (defaults to https://app.parapet.run if you omit the -e override).
serve (no other flags) runs the MCP server over stdio — this is what an MCP client actually launches; you don't run it by hand.

Once connected, ask Claude Code to build a demo, add Parapet to an existing MAF/ADK project, or audit an existing codebase for ungoverned model/tool calls, and the relevant skill takes it from there — see Skills for what each one does, and MCP tools for the individual tools they call along the way.

Environment¶

Variable	Default	Controls
`PARAPETAI_CONTROL_PLANE_URL`	`https://app.parapet.run`	The control plane every `parapet_*` tool talks to, unless a per-call argument overrides it
`PARAPETAI_MCP_CONFIG_DIR`	`~/.parapet`	Where the CLI token (`credentials.json`, one entry per control-plane URL) is stored after `parapet_login_start`/`parapet_login_wait`

Never paste a token, agent secret, or any other credential into chat — the login flow is designed specifically to avoid that: parapet_login_start opens a browser approval page (and hands back a manual fallback URL/code either way), parapet_login_wait polls until you approve, and neither ever returns the credential itself to the calling model.

Source¶

MIT licensed, part of the same repo as the SDK: github.com/Parapet-run/parapet-agenticai-sdk/tree/main/mcp-server.