pullboard-dev/pullboard

Autonomous development for real products. You set the standard. Declare your vision. Walk away. It builds.

What it solves

When you let AI agents write code, they often say "done" when they aren't, forget your earlier decisions, and step on each other's work. Without enforcement, you end up reviewing every line and repeating yourself. Pullboard fixes this by making your intent explicit and by having a second agent independently verify every change before it counts.

How it works

Pullboard gives agents a shared plan and separate lanes, and nothing is accepted without proof. You approve rows in a spec file (SPEC.md), each with a clear requirement and a test that proves it. Each agent works in its own isolated worktree so they never edit the same file. A gate command (like npm test or pytest) must pass at the exact commit an agent submits. Then a different agent—a verifier—checks that precise commit against the spec. If it passes, the work is accepted and recorded; if not, it's sent back with the reason.

Agents talk through "shouts": questions go up a chain, and only the decisions that need you are passed along. A coordinator handles what it can, so you only get the calls that truly need your judgement. All rules live in DOCTRINE.md, and a full activity log records every claim, submit, reject, and accept.

Who it's for

Anyone using AI coding agents—like Claude Code, Codex, or any MCP client—on a serious software project. It's built for both solo developers and teams running dozens of agents over months, where keeping the project coherent is the whole battle.

Highlights

  • Spec-driven development: Only your approved requirements count, and each item freezes its acceptance bar when an agent claims it.
  • Isolated lanes: Every agent works in its own worktree, so they can't collide on the same file.
  • Independent verification: A second agent checks the exact commit, not just the claim, so "done" is proven.
  • Your judgement is the bar: You approve rows, answer the few questions that reach you, and the agents fill in the blanks.
  • Runs entirely on your machine: Everything lives in your repo and local SQLite—no hosted service, no dependencies.
  • Works with your tools: Plug in Claude Code, Codex, or any MCP client via simple commands.
  • Blueprints for common stacks: Prebuilt spec and rule templates for web apps, libraries, iOS apps, data pipelines, and more.

Related

  • Project
  • Project
  • Project
  • Project