Mecha

Managing event-driven agentic workflows — one binary, YAML config, policy-controlled.

Run Claude, Codex, or Gemini on every PR

GitHub or GitLab webhook arrives, mecha matches it to a worker, dispatches the prompt, writes the result back. Automatic.

Switch models in one line

Docker workers for Claude/Codex/Gemini. Adapters for Ollama, vLLM, or any OpenAI-compatible API. Same contract, swap the YAML.

Control what AI can write back

Policy decides per-worker — allow comments, block labels, limit length, deny commits. One YAML block. One security boundary.

Self-hosted single binary

Built in Go. No cloud dependency. No runtime. Your code stays on your infra. Works on macOS and Linux.

Define a worker. Start it. Done.

yaml

name: pr-reviewer
docker:
  image: mecha-worker:latest
  token: claude.default
  env:
    CLAUDE_MODEL: claude-sonnet-4-6
events:
  - source: github
    on: [pull_request.opened]
    prompt: "Review this PR for security issues.\n\n{{.diff}}"
policy:
  comment: { allow: true }
  status: { allow: true }
  commit: { allow: false }

bash

mecha worker add workers/pr-reviewer.yml
mecha serve

Every PR now gets an automated review — comments, status checks, labels — controlled by policy.

The pipeline

Event.arrive → Event.match → Task.create → Task.dispatch → Policy.filter → Task.complete

Four nouns: Event, Worker, Task, Policy. The pipeline is deterministic. The LLM is the only smart part. Policy is the only gate.