A board where agents check each other's homework before you see it

What it does

Pullboard is a coordination layer that lives in your git repo and runs a fleet of coding agents (Claude Code, Codex, any MCP client) against a spec you approve. Requirements are rows in SPEC.md, each with a gate that proves it; rules live in DOCTRINE.md, enforced by everything from git hooks to review. Each agent works in its own worktree and lane, and nothing merges until a different agent verifies that exact commit against the frozen requirement.

The interesting bit

The boring part is the value: it’s a queue with receipts. Every claim, submit, reject and accept is on the record, and on Pullboard’s own board roughly one in four “done” claims got sent back by a verifier (227 of 908). The decision chain is also deliberate — agents ask their coordinator first, and only genuinely human judgement reaches you, which is what makes walking away plausible.

Key highlights

• Local-first, no dependencies, MIT; the board runs on a local API your own scripts can hit
• Blueprints: bundled spec-rule starter kits for web apps, libraries, iOS apps, data pipelines and more
• A desktop/phone view showing only the decisions that need you, plus full review history per item
• Dogfooded hard: 490 agents cycled through their own board; one overnight run landed 11 items to main with no human moves
• view --export writes a static, replayable snapshot of a board

Caveats

• The impressive numbers (hundreds of agents, overnight factories) are all from Pullboard’s own board — your mileage on a fresh project is unproven in the README
• The methodology has real vocabulary to learn (items, shouts, doctrine, receipts); it’s a workflow commitment, not a drop-in
• Still pre-1.0 (v0.9.3), so expect churn

Verdict

Worth a look if you’re already running multiple coding agents and drowning in chat logs and unreviewed diffs. If one agent in one terminal works for you, this is machinery you don’t need yet.