agentsitter

Specs in, PRs out.

A self-healing pipeline that turns specs into reviewed, tested pull requests.

Your AI assistant is the manager. Coding agents are the team. agentsitter is the infrastructure between them.

Install

curl -fsSL https://agentsitter.dev/install.sh | sh

Linux only (amd64 and arm64). macOS support is planned.

Build from source

git clone https://github.com/agentsitter/agentsitter.git
cd agentsitter
go build -o sitter ./cmd/agentsitter
sudo mv sitter /usr/local/bin/

Quick start

1. Start the server

sitter serve

2. In a separate terminal, create a work item from a spec

sitter create --title "Add dark mode" --context @spec.md

3. Launch a coding agent

sitter start add-dark-mode

4. Review the PR when it's ready

The agent clones your repo into an isolated worktree, implements the feature, passes automated review, creates a PR, and fixes CI failures — all without intervention.

How it works

1. Create — Your AI assistant creates a work item with a spec
2. Code — A coding agent is spawned in an isolated worktree to implement the spec
3. Review — A review agent inspects every line against the spec. Issues found? The coder is restarted with feedback. No warnings pass. No rubber stamps
4. Deliver — Branch pushed, PR created, labels synced. CI fails? The coding agent fixes it

Retry limits per failure type (up to 3 CI attempts, up to 5 review attempts by default) — no infinite loops. Review retries are configurable.

Features

• Agent-first orchestration — CLI and REST API — your AI picks whichever fits. Create work items, start agents, and monitor progress from the terminal or over HTTP
• Self-healing pipeline — CI fails? Agent restarts with the error (up to 3 attempts). Review rejects? Feedback loop (up to 5 attempts by default, configurable). Agent hangs? Detected and restarted
• Strict automated review — A review agent checks every PR against the spec. Zero warnings policy — both blocking issues and warnings must be resolved
• Work item queue with dependencies — Define specs, set blocked-by dependencies, tag by type. Items auto-start when slots and dependencies clear
• Observable by design — Activity logs, status polling, and SSE event streams. Your orchestrating agent always knows what every coding agent is doing
• Single binary, self-hosted — One binary. SQLite. Under 20MB. No cloud dependencies, no Docker, no infrastructure. Your code never leaves your machine

Supported agents

• Claude Code (claude) — supported
• Codex, OpenCode, Gemini CLI, Cursor Agent — coming soon

Documentation

Full documentation at agentsitter.dev/docs.

• Getting started
• CLI reference
• GitHub App setup
• API reference

License

MIT