Skip to content
PR Bot
Menu

MCP server

The PR Bot MCP server

Let an agent trigger a review, check a review’s status, list reviewer bypasses and read PR stats, with dry runs as the default.

See all toolsMCP clients

Two ways to connect

The API can serve MCP itself on POST /mcp, a stateless Streamable HTTP route. Or an MCP client such as Claude Code can launch @pr-bot/mcp-local, a stdio server that calls a running PR Bot API over HTTP. Both offer the same four tools.

Off until you turn it on

The /mcp route is disabled by default and appears only when PRBOT_MCP_SERVER_ENABLED is true. Its API key and allowed origins live in the database and are managed in the web app under Settings, MCP Server; a new key is shown once.

The key is an exposure control, not an authorization boundary: the PR Bot API has no inbound authentication of its own, so keep it on a network you trust.

Dry runs by default

An agent skips the chat UI’s human confirmation step, so the MCP trigger_review tool does a dry run unless it is told otherwise: the full review runs and nothing is posted to the PR.

Not the MCP bridge

The MCP server is PR Bot answering agents. The MCP bridge is the opposite direction: PR Bot calling the separate mcp-agent-bridge service for its own AI work. The two are unrelated.

Shell and skill

The same operations are a CLI (pr-bot, built and linked from the repo with make link-cli) and an agent skill, skills/pr-bot/SKILL.md, for agents that have a shell but no MCP client.

MCP tools