parapetai-mcp

MCP server for Claude Code (and other MCP clients) that talks to a Parapet control plane: device-code login in your browser, then provision governed agents and pull the exact quickstart config a codebase needs — without hand-editing env vars.

pipx install parapetai-mcp
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

pipx, not pip — this is 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" (Homebrew's Python does), which makes a bare pip install parapetai-mcp fail rather than risk corrupting packages the system relies on. pipx installs CLI tools into their own isolated venv and puts the entry point on PATH — no venv to activate, no --break-system-packages flag needed. No pipx available? python3 -m pip install --user pipx && pipx ensurepath (or brew install pipx on macOS), then re-run the command above. uv tool install parapetai-mcp is an equivalent one-liner if uv is already on the machine.

Then in Claude Code: "add Parapet to this project."

Tools

A parapet_getting_started prompt is also registered — clients that list MCP prompts (e.g. Claude Code's /mcp menu) can pick it as a first-run menu: build an example governed-agent app, or do something else with the tools above. It's discoverable, not auto-fired on connect — no MCP mechanism forces a prompt onto the first turn.

This server itself never edits your project's source files — it only talks to the control plane, with two exceptions that stay purely local: parapet_check_prerequisites only ever reads local machine state, never sending anything anywhere, and parapet_audit_codebase only ever reads your source and writes one report file (.parapet/audit/report.md by default) — actually instrumenting flagged code is a separate, explicit step (the parapet-audit-fix skill, driven by Claude Code's own file tools, not this server). Seven packaged skills tell Claude Code how to use these tools and where to write the resulting config — parapet-maf for a Microsoft Agent Framework project, parapet-adk for a Google ADK one, parapet-langgraph for a LangGraph/LangChain one (the frameworks put governance in different places, so the instrumentation steps genuinely differ; Claude Code picks whichever matches your project), parapet-quickdemo for generating a runnable governed-vs-ungoverned example from scratch in a new or empty project (templates for all three frameworks), parapet-install-prereqs for detecting and (with your per-step approval) installing Python/pipx/uv when one of the others needs them, and parapet-audit / parapet-audit-fix for scanning an existing codebase for ungoverned model/tool calls and then, as a separate step, wrapping the flagged sites in GovernedAgent/GovernedRunner.

Point at a different control plane (e.g. a local make dev instance) with PARAPETAI_CONTROL_PLANE_URL (default https://app.parapet.run, the hosted control plane).