Pullboard runs a team of agents on your repo for days. They claim the work, build it, check each other and merge, all held to your spec and your rules.
One SQLite file in your .git. No account, no server.
Run pullboard view and one browser tab shows every task, which agent has it, and anything waiting on you.
A second agent tests every commit before it merges, so bugs go back with a reason and only verified work lands.
3f9a1cThe builder can't verify its own work
web-1 fixes it and submits again
Item #2 is verified, with both verdicts on record
Git hooks hold every commit to your doctrine, and only the spec rows you approve get built.
A claim gives each item one agent, and each agent owns its lane's folders in its own worktree.
The first claim wins
HELDweb-2 moves on to #4 Cart total with tax. A stalled lease lapses.
When an item’s prerequisites pass, the next agent claims it unprompted, and resume picks up any session.
$ pullboard resume
resume: web-1, web lane, at ~/storefront-web-1
branch web/1: 0 ahead of main, 0 behind
2 unread shouts; newest from api-1: Region only for now;
postcode lands with the next item.
next: pullboard next (3 ready in your lane)
$ Six more things pullboard init sets up for you.
One line to a lane, an agent or everyone, and your answer flows back down.
Pullboard's standard rules plus your own, checked by git hooks at every commit.
A submit is refused unless your tests pass at that exact commit.
Route simple items to a cheaper model and keep hard ones on a strong one.
Claude Code, Codex, OpenCode or anything that runs in a shell.
Each accept or reject records who checked, which commit and why.
Agents talking, and a decision that went from web-1 to the coordinator, on to you, and back.
The house rules from DOCTRINE.md, the standard set every repo starts with, each one approved or still a draft.
review-1 sent #1 back with a reason, then accepted the fix. Both verdicts stay on the task, each at the commit it checked.
api-1 holds #5. Another agent who tries to claim it is refused, and the claim lapses if its lease runs out.
Task #6 waits on #2. Nobody can claim it until #2 is verified, and then it is next in line.
$ pullboard submit 1
pullboard: [GATE_RED] the submission proof is red at b4d22aa36834; fix it,
commit, submit again. gate red in 0s:
not ok 1 - pays by card in one step
error: "'two steps' == 'one step'"
$ A real refusal: the test fails at that commit, so the submit is refused until it passes.
$ pullboard worktree web
made shop-web-2 on branch web/2, joined as web-2 in the web lane
cd shop-web-2 && pullboard inbox
cd shop-web-2 && pullboard next
$ Any agent that runs a shell joins the same way: a worktree, a lane and a name.
Each agent shows its lane and route: strong agents take any item, light ones only items routed light.
You own the doctrine and spec. Work flows down to the agents, questions come back up, and only your calls reach you.
Not every job needs a board. Take the fit test.
Install it, run pullboard init in your repo, and tell your agent what to build.