Harness — Claude Code that ships tasks unattended

A kanban board where a two-line draft becomes a pull request: Claude Code agents write the spec, ask questions, plan, prototype, build, test in a real browser and judge the result.

The board: tasks from draft to delivered, with cost per task

problem: AI coding agents are fast but unreliable on their own: they skip steps, drift from the request and call it done too early — so a developer still babysits every task.
built: The whole system, solo (23 of 23 commits, plus 16 of 16 in the guide site): a zero-dependency Node core that runs one Claude Code process per unit of work; contracts in code, not instructions in a prompt, that gate every step; QA and designer agents that drive a real browser through Playwright; a judge that can reject a round; delivery as GitHub pull requests; a Next.js dashboard with live agent traces over SSE, cost per stage and task, and Telegram notifications. Plus harness-info, a bilingual guide that teaches the approach.
result: 4 → 4pilot tasks → pull requests$9.13API cost of the whole pilot47units of agent work, each gated

Claude CodeNode.jsTypeScriptNext.js 16HeroUITailwind CSSPlaywrightSSEGitHub

Live site ↗Source ↗

Live trace: which agent ran, which tools it used, what came out

Stats: time, tokens and cost by stage and by task

Need something like this? Let’s talk.

Start a project