Shipwright Harness -- autonomous delivery agent for Claude Code
The open-source autonomous delivery agent for Claude Code. A deployable cloud agent and the autonomous coding system that powers it — built on the Shipwright plugin, running on your own codebase.
It builds itself, and the numbers are public
Shipwright's own delivery pipeline runs on Shipwright. Cycle time, CI first-pass rate,
review verdicts and estimation accuracy for this repository are public:
→ proof.shipwrightharness.com
As of 2026-09-25, trailing 30 days: 336 tasks completed, 60 PRs merged, 85.7% CI
first-pass rate, ~4.4h median cycle time. No other autonomous coding agent publishes its
own delivery metrics — you can check our work before installing anything.
What is Shipwright Harness?
Two faces, one product:
- The agent — deploy it to your cloud (GitHub Actions or self-hosted). It does autonomous coding on your codebase, held to the same review and test bar as human code.
- The system — the autonomous coding system, built on the Claude Code
shipwright plugin: plan · build · review · metrics. Use it interactively inside Claude Code, or let the agent run it autonomously.
It runs in your environment, on your codebase — you own it, it's MIT, and it's free.
What it does
Shipwright turns a feature idea into shipped, reviewed code through a sequence of Claude Code commands — each stage producing a durable artifact the next stage consumes:
- Write a PRD for your idea — a structured product spec ready for /plan-session.
- Plan the spec into a queue of well-scoped, dependency-ordered tasks (tracked in the Shipwright task store — a shared queue any agent or contributor can query).
- Execute the next ready task — build, test, and open a PR.
- Review the PR with policy-controlled, inline feedback.
- Ship the merged change.
Install
/plugin install shipwright@app-vitals/shipwright
Requires Claude Code. Point it at your own repository — Shipwright is repo-agnostic.
⭐ If the approach makes sense to you, star the repo — it's how we gauge whether to keep
investing in the open-source track.
Deploying the services to Kubernetes? The shipwright Helm chart is published to a Helm repo on each chart version bump:
helm repo add shipwright https://app-vitals.github.io/shipwright
helm install my-release shipwright/shipwright --namespace shipwright --create-namespace
See docs/deploy-kubernetes.md for the Kubernetes deployment guide and docs/deploy-kubernetes-providers.md for end-to-end per-provider walkthroughs (Minikube / GKE / EKS / Traefik / cloud-native), and docs/helm-repo.md for the published-repo flow and how publishing is triggered.
Get Shipwright in your Slack
The Shipwright agent runs as a Slack app — DMs, @mentions, and cron notifications, all in Socket Mode. Two ways to get there:
Quickstart
Run the metrics dashboard locally today — offline by default: fixtures-backed, no accounts, no database. One copy-paste prompt covers both steps (terminal + an in-session slash command) and opens the dashboard.
Paste this into a Claude Code session:
Set up Shipwright Harness locally and open the metrics dashboard.
1. In a terminal, run:
git clone https://github.com/app-vitals/shipwright.git && cd shipwright && ./scripts/quickstart.sh
Checks prerequisites, installs dependencies (task setup), and starts the
metrics dashboard offline (no accounts or secrets needed). Leave it running.
2. Inside this Claude Code session, install the plugin:
/plugin install shipwright@app-vitals/shipwright
3. Open the dashboard in your browser:
http://localhost:3460/dashboard
Step 1 runs in your terminal; step 2 runs inside the Claude Code session.
Prerequisites: Claude Code, git, Bun, and go-task. Full details, the QUICKSTART_SKIP_SERVE CI guard, and the offline-default explanation live in docs/quickstart.md.
Full dev stack
The complete local stack — metrics dashboard, admin UI, task-store, and the Shipwright agent in Docker — needs tmux, Docker, and a local PostgreSQL instance (port 5432). Paste this into a Claude Code session:
Set up the full Shipwright Harness dev stack locally.
Prerequisites: tmux, Docker, PostgreSQL running on localhost:5432, Bun, go-task.
1. In a terminal, clone and set up:
git clone https://github.com/app-vitals/shipwright.git && cd shipwright && task setup
2. Launch the full stack (6-pane tmux session):
task stack
First run: task stack auto-creates state/dev-agent.env from the example
and exits — open it and set one of these, then re-run task stack:
CLAUDE_CODE_OAUTH_TOKEN=<your token> (run: claude /oauth-token)
ANTHROPIC_API_KEY=<your key> (https://console.anthropic.com/ → API Keys)
Opens a tmux session named "shipwright" with 6 panes:
metrics (:3460) admin (:3001) task-store (:3002) chat-svc (:3003) agent logs
3. Open the dashboard in your browser:
http://localhost:3460/dashboard
4. Inside a Claude Code session:
/plugin install shipwright@app-vitals/shipwright
To stop: tmux kill-session -t shipwright
Why Shipwright Harness
- Free and open-source (MIT) — the own-it alternative to closed, hosted coding agents. No rented infrastructure, no lock-in.
- Runs in your environment, your cloud — your code never leaves your control.
- The same quality bar as human code — tests land with the code, gated by a five-phase test-readiness pipeline, so an autonomous agent can be trusted.
- Metrics on your own pipeline — first-time-quality rate, estimation accuracy, and review-verdict trends, measured on your delivery.
- Built on Claude Code — we use it every day, and Shipwright extends it rather than replacing it.
Components
| Component | What it does | Status |
|---|
| Plugin (the system) | The shipwright toolchain you /plugin install — planning, queue-based execution, review, a test-readiness pipeline, and deploy commands. | ✅ Available |
| Metrics dashboard | A stateless service that reads pipeline analytics (task throughput, CI first-pass rate, review verdicts, estimation accuracy) and renders a dashboard. Run locally with task api or task ui (offline mode, no secrets needed). | ✅ Available |
| Shipwright agent | A thin autonomous runner that drives the system on a schedule — pick the next ready task → build → ship a PR → forward metrics — deployable to GitHub Actions or self-hosted. | ✅ Available |
The workflow
/shipwright:prd → a product spec
/shipwright:plan-session → a dependency-ordered task queue
/shipwright:dev-task → build + test + open a PR for the next ready task
/shipwright:review → policy-controlled PR review
/shipwright:patch → address review findings / failing CI
/shipwright:deploy → merge + deploy
Tasks are tracked in the Shipwright task store (see docs/task-store.md), so the queue is shared across every agent working the pipeline.
Project status
See it live — Shipwright builds itself. The public metrics dashboard shows live pipeline data for this repo: first-time quality, cycle time, estimation accuracy, and task throughput — all generated by the same agent and plugin you're reading about.
Shipwright Harness is live — plugin, metrics dashboard, and the Shipwright agent all ship together (see the release badge above for the current version). See the issues for the live roadmap and upcoming improvements.
The metrics dashboard is runnable locally today — the Quickstart wraps this in one copy-paste prompt (./scripts/quickstart.sh). The underlying tasks:
task setup # bun install
task api # start metrics dashboard in offline mode → http://localhost:3460/dashboard
task dev # dev supervisor: starts metrics + Ctrl-C kills all children
task stack # full dev stack in a tmux session (6 panes) — requires tmux
task stack brings up a single tmux session (shipwright) with a 6-pane dashboard: metrics (SQLite, :3460), admin (CRUD API + UI, :3001), task-store (:3002), chat-svc (the platform-hosted chat service from app-vitals/agent, :3003 — see local platform development), the agent in Docker with the chat poll loop enabled, and a scratch logs shell. Chatting with the agent happens in the browser via the admin console's Chat tab (/admin/chat), not a REPL or an endpoint on the agent itself. It runs a Prisma migrate deploy preflight first so the admin service's Postgres schema is up to date; on macOS the preflight checks Postgres is reachable and offers to run the needed brew/createdb commands if it isn't. task stack requires tmux; if it isn't installed, use task dev (the no-tmux fallback that starts the metrics dashboard).
See docs/quickstart.md for the full onboarding prompt and offline-default behavior.
Built on Claude Code
Shipwright Harness is a Claude Code plugin through and through — built on it, for it, and used with it daily. If you already run Claude Code, Shipwright is a /plugin install away.
See it in the wild
- Compare it — how Shipwright's plan → queue → loop model differs from the commercial agent tier and from other Claude Code orchestrators: see the comparison pages.
- Have us run it on your pipeline — the design-partner program: the people who built Shipwright, hands-on on your codebase.
- Read the full story — shipwrightharness.com.
Test system
Shipwright enforces a four-layer test architecture (unit / integration / smoke / e2e) across all three components. Layer boundaries, per-component run commands, speed budgets, and the test-isolation contract are defined in docs/test-readiness/test-system.md.
Configuration
All configuration options — plugin env vars, agent env vars, and policy fields — are documented in docs/configuration.md.
Observability
Optional Sentry error/log reporting, disabled by default and self-host-friendly. See docs/observability.md for exactly what is collected.
Contributing
Issues and discussion are welcome. See CONTRIBUTING.md for conventions and workflow, and our CODE_OF_CONDUCT.md. This repository is MIT-licensed and public — please keep contributions free of any proprietary or confidential material.
License
MIT © 2026 App Vitals
Naming
Brand vs. package: the project is Shipwright Harness; the plugin/package you install is shipwright.
