Herdkit

herdkit

**Turn your repo into an autonomous development shop.** You talk to one coordinator agent in plain language — it files the work, spawns a herd of AI builders in parallel, gates every merge behind adversarial review, and heals its own failures. You come back to merged PRs and a truthful console, not a stalled session.

Agent SDKs help you *build* agents. Coding agents *perform tasks*. Merge queues *gate humans*. herdkit runs the governed development operation that connects all three — locally, with any model, with receipts.

you › drain the backlog overnight

coordinator › 14 open items. Mapped which items touch the same files — spawning 5
              builders in parallel, holding 2 that collide until their blocker merges.
              Reviews run up to 8 in parallel. I'll route failures and report at milestones.

coordinator › Morning report: 11 PRs merged (avg $9.62/PR — ledger below), 2 in
              review, 1 escalation: #907's conflict is semantic — two rewrites of
              the same function — that one's your call.
              Replay any merge's full gate history:  herd why 903

Works with **Claude Code, OpenAI Codex, Grok, and Cursor** — mixed freely, in one control room. Local control plane: your machine, your repo, your keys. No hosted service.

**The receipts:** herdkit is built by itself. **596 merged PRs** on this repo were filed, built, adversarially reviewed, and merged by the herd — **$9.62 per merged PR** on the recorded ledger, with the reviewer bouncing **39%** of PRs back with findings before they merged. Run `herd why ` on any of them — or read [a committed replay of one real merge](docs/assets/sample-why-replay.txt) right now.

---

Why this needs to exist

If you've used a coding agent for real work, you know the four walls:

• **Babysit or s