ShipCode

AI Agents that ship working code. GitHub issue in, reviewed PR out.

Warning

This project is under active development and is a work in progress. Features may be incomplete, APIs may change, and there may be bugs. Contributions and feedback welcome!

Watch the launch video (MP4)

Label a GitHub issue with shipcode:agent:claude, shipcode:agent:codex, shipcode:agent:openrouter, shipcode:agent:openrouter/auto, or shipcode:agent:openrouter/free, and ShipCode handles the rest: plan, adversarial review, implement, verify, and ship a PR. Approval gates are configurable globally, per project, and per issue.

Install

brew install --cask shipshitdev/tap/shipcode   # macOS desktop app
npx @shipshitdev/shipcode                      # CLI

macOS users: the v0.1.0 build is unsigned. After install, run xattr -cr "/Applications/ShipCode.app" once to clear the Gatekeeper quarantine.

How it works

GitHub Issue (labeled shipcode:agent:claude / codex / openrouter)
       |
       v
    PLAN (Opus 4.6) --- rewrite issue into spec + structured plan
       |
       v
    REVIEW (configured model) --- adversarial critique
       |
       +-- APPROVE --> EXECUTE
       +-- REVISE (max 2 rounds) --> re-review
       +-- REJECT --> FAILED
       |
       v
    EXECUTE (routed model) --- implement in isolated git worktree
       |
       v
    VERIFY (configured model) --- test, runtime QA, check diff against acceptance criteria
       |
       v
    SHIP --- push branch, create PR, link to issue

Architecture

Turborepo monorepo with Electron desktop app:

apps/
  web/              Marketing site (Next.js + Tailwind)
  desktop/          Electron + Vite + React 19
  cli/              CLI tool (published as '@shipshitdev/shipcode' on npm)
  docs/             Documentation (Nextra)
packages/
  pipeline/         Pipeline state machine
  shared/           Types, schemas, constants
  agents/           Process manager, prompts, GitHub CLI wrapper
  db/               SQLite persistence (node:sqlite)
  git/              Git operations + worktree manager
  ui/               React components (kanban board, viewers)

Prerequisites

• Bun >= 1.2
• Claude Code CLI (claude)
• Codex CLI (codex)
• GitHub CLI (gh) with auth configured
• Git

Install

Desktop app (macOS, Homebrew):

brew tap shipshitdev/tap
brew install --cask shipcode

macOS warning: the v0.1.0 desktop build is unsigned and not notarized yet. If macOS blocks the app after Homebrew installs it, run:

xattr -cr "/Applications/ShipCode.app"

This is a temporary pre-release workaround until ShipCode ships with Developer ID signing and notarization.

CLI (Node.js >= 22.5.0):

npx @shipshitdev/shipcode onboard
npx @shipshitdev/shipcode run 42

# Optional global install (exposes `shipcode` on $PATH)
npm install -g @shipshitdev/shipcode

From source:

git clone https://github.com/shipshitdev/shipcode
cd shipcode
bun install
bun run dev:desktop

GitHub integration

ShipCode uses the gh CLI for all GitHub operations. Make sure you're authenticated:

gh auth status

Labels

Label	Effect
shipcode:agent:claude	Claude implements the issue
shipcode:agent:codex	Codex implements the issue
shipcode:agent:openrouter	OpenRouter implements the issue with the configured executor model
shipcode:agent:openrouter/auto	OpenRouter chooses the model through its auto router
shipcode:agent:openrouter/free	OpenRouter stays on the configured free-tier model

Pipeline state labels (set automatically)

Label	Meaning
shipcode:pipeline:queued	Issue picked up, waiting for pipeline slot
shipcode:pipeline:planning	Planner is generating a plan
shipcode:pipeline:reviewing	Reviewer is critiquing the plan
shipcode:pipeline:executing	Executor is implementing changes
shipcode:pipeline:testing	Test/runtime QA commands are running
shipcode:pipeline:verifying	Verifier is checking the result
shipcode:pipeline:shipping	Branch push and PR creation are running
shipcode:pipeline:failed	Pipeline failed

Key design decisions

• gh CLI over GitHub API -- no OAuth, no token management, uses existing auth
• Adversarial review -- configurable planner/reviewer models catch different blind spots before execution
• Configurable review rounds -- default is fast-path zero revisions, with project and issue overrides up to 5
• Optional approval gates -- stay fully autonomous by default, or pause before execution when the task needs a human checkpoint
• Fork-point SHA for diffs -- stable diffs even when base branch moves

License

MIT