ShipCode

Autonomous AI coding pipeline. GitHub issues in, pull requests out.

Label a GitHub issue with shipcode:agent:claude, shipcode:agent:codex, shipcode:agent:openrouter, shipcode:agent:openrouter/auto, or shipcode:agent:openrouter/free. ShipCode does the rest — plan, adversarial review, implement, test/runtime-QA, verify, ship a PR — with optional approval gates when you want a human checkpoint before execution.

Install

ShipCode ships as two separate products that share the same pipeline engine. Most users only need one of them.

Desktop app (macOS) — recommended if you want the GUI:


brew tap shipshitdev/tap
brew install --cask shipcode

CLI — for terminal, CI, cron jobs, and headless environments:


# Run once without installing
npx @shipshitdev/shipcode onboard
 
# Or install globally
npm install -g @shipshitdev/shipcode
shipcode onboard
shipcode run 42

See Getting Started for prerequisites and CLI vs Desktop for which one fits your workflow.

How it works

Phase	Default model	What happens
Plan	Opus 4.6	Rewrites the issue into a spec and structured plan
Review	Codex (high reasoning) — set during onboarding	Adversarial critique, configurable revision rounds
Execute	Routed per label (claude / codex / openrouter) — configurable in Settings	Implements in an isolated git worktree
Test / QA	Repo setup contract	Runs tests, runtime QA, and optional visual QA
Verify	Opus 4.6 — set during onboarding	Checks the diff, test output, and acceptance criteria
Ship	—	Pushes branch, creates PR, links to issue

Why ShipCode?

Most AI coding tools require you to stay in the loop. ShipCode runs the full pipeline unattended — planning through PR creation — while you work on something else.

Adversarial review catches bad plans before a single line is written
Isolated worktrees mean each pipeline run is sandboxed from your working tree
Routed models let you mix Claude, Codex, and OpenRouter models per phase
Desktop operations views cover automations, worktree cleanup, PR triage, project insights, and cost tracking