Specpipe

One spec-first workflow, installed into every AI coding agent you use.

Live demo →

Specpipe installs a disciplined loop — **spec → code + tests → build pass** — as native skills, guard hooks, and project rules. You author it once; it lands in whichever agent you run (Claude Code, Codex, Cursor, OpenCode, Antigravity, OpenClaw, Hermes). Switch agents, keep the same `/sp-*` commands.

npx specpipe init        # interactive picker — choose agents, skills, guards

---

Why

AI agents write code fast and drift fast — they invent requirements, test the wrong things, and quietly break what worked. Specpipe fixes the loop, not the model:

• The spec is the source of truth. Acceptance scenarios (Given/When/Then) live inside the spec; if code contradicts it, the code is wrong. No separate test plan to fall out of sync.
• Tests check behavior, not guesses. They come from the spec, so they survive refactors.
• Guardrails are enforced, not suggested. Hooks stop an agent from reading secrets or crawling node_modules before it happens.
• Learn it once. The same commands and discipline work in every agent.

You don't have to be a developer to feel it: describe a feature in plain language, the agent turns it into a spec you can read, builds against it, and proves it with tests — every step reviewable.

---

Quick start

npx specpipe init        # interactive: scope (project/global), agents, skills, guards
npx specpipe init -y     # non-interactive defaults: this project, Claude, all skills + guards

Then open your agent (`claude`, `codex`, `cursor`, …) and run the loop. The transcript below is on Claude Code with a Swift project; the same commands and output work in