The MCP server

AI coding agents write infrastructure code now. garboard mcp puts the same deterministic gate that reviews your pull requests inside the agent's loop, over the Model Context Protocol — so an agent can learn your repo's conventions and check its own diff before a PR ever exists.

Setup

Scan the repo once, then register the server with your agent. With Claude Code:

garboard scan            # writes .garboard/rcd.json — works offline, no API key
claude mcp add garboard -- garboard mcp

The server reads the Repository Convention Document produced by the scan. Everything runs locally over stdio: no network, no credentials, and it works with no Claude API key configured — the gate is deterministic, not AI.

Tools exposed

get_conventions — the repo's derived conventions with file:line evidence, so the agent writes changes that match the codebase instead of generic Terraform.

check_diff — runs the review gate over changed Terraform/OpenTofu files and returns findings. This is the same check external pull requests get; an agent that calls it before opening a PR arrives clean.

get_exemplars — the repo's best existing example of a resource type, as file references and snippets, for the agent to imitate.

search_episodes — searches remembered merge history, so past review decisions inform new changes.

Contract

The MCP surface is read-only — it never mutates the repo, the database, or any cloud. A missing .garboard/rcd.json degrades gracefully: check_diff still works; the convention tools return empty until you scan. Available on every plan.

Next: self-hosting →

See it on your repo

The deterministic gate runs free on a real repository, or we can walk you through it live.