What is Pullboard?
Pullboard is a command-line coordinator for teams of coding agents. You write requirements as rows in SPEC.md and house rules in DOCTRINE.md; agents claim items from a board, build them in their own git worktrees, and submit commits that a separate verifier agent must accept before they land on main. A coordinator agent fields questions and passes up only the decisions that need you.
The board is one SQLite file inside .git, with no hosted service required. Each lane owns a set of paths and spec sections, and a gate command such as npm test or pytest runs before every submit and push. A local web view shows every project on the machine, with the claims, rejections and decisions waiting for you. An opt-in relay at pullboard.dev puts the board on your phone and other machines, with board contents encrypted.
It is written in JavaScript, needs Node.js 22.13 or later, installs with npm i -g pullboard, and is MIT licensed. Claude Code gets bundled skills; Codex and other agents read AGENTS.md or a role prompt.
Is Pullboard worth it?
Pullboard's useful idea is that an agent's claim to be done counts for nothing until another agent has checked out the exact commit and tried to break it. Most multi-agent setups give you parallel builders; Pullboard spends as much effort on the verifier, the frozen acceptance criterion and the audit trail as on the building.
The project publishes numbers from its own development, which is more than most tools in this space do. Between 4 and 16 October 2026, according to the README, verifiers sent work back 227 times out of 908 submissions, 588 items merged and none merged without an accept. One night on 16 October ran 38 agents for 9.6 hours with no human moves, landing 11 items, 6 features and 5 fixes, with 1,596 lines of tests against 817 lines of product code. These are the maintainer's own figures from building Pullboard with Pullboard, not an independent benchmark.
The cost is ceremony. Every requirement becomes a SPEC.md row with a gate, every lane gets owned paths in pullboard.json, and agents must follow the claim, submit and verify loop. That suits a long-running product with a test suite; it is heavy for a weekend script. Ten bundled blueprints for web apps, libraries, iOS apps and data pipelines shorten the start, and pullboard tour shows the loop on a throwaway repository.
The guards are softer than they look. The README says worktree identity is not a security boundary, git hooks can be skipped with --no-verify, and Windows is not tested in CI. The repository was created on 7 October 2026, reached v0.9.3 on 27 October, and has 690 stars, with one author named in its npm package. The relay keeps board contents sealed with AES-256-GCM, but its architecture document says it can read repository names, timing and the exact command text of phone approvals, which it keeps with no deletion path.
Against Vibe Kanban, which focuses on running agents and reviewing their diffs yourself, Pullboard tries to take you out of line-by-line review entirely and replace it with an adversarial agent and a written spec. Whether that holds depends on how good your tests are, since the gate and the verifier both lean on them.
Where it earns its place
A solo founder with a tested Node.js web app
Let agents work overnight without waking up to broken main
Approve SPEC.md rows with a test as each gate, start a coordinator in Claude Code, and add builder and verifier worktrees. Rejected work goes back with a reason; only accepted commits merge. Cost is the model usage of every agent, which is not small with several running for hours.
A three-person agency maintaining client sites
Stop re-explaining the same house rules to every agent session
Write the rules once in DOCTRINE.md and save the setup as a blueprint with pullboard blueprint save, then start each new client repository from it. Hooks enforce what a machine can check; the verifier covers the rest.
An engineering lead piloting autonomous agents
Show stakeholders an audit trail of what agents did and why
Every claim, submit, rejection and accept is recorded with a reason, and pullboard view --export writes a static, replayable snapshot to share. The trail is only as convincing as the tests behind each gate.
Before you commit
- Setup
- npm i -g pullboard on Node.js 22.13 or later, then pullboard init in a git repository and commit the files it writes; pullboard tour runs the full loop on a throwaway repo first.
- Integrations
- Claude Code skills, AGENTS.md and role prompts for Codex and others, MCP clients, a local JSON API (v1) behind the view, and any test command as the gate: npm test, pytest, go test, cargo test and more.
- Data and privacy
- Board is a SQLite file in .git; spec, doctrine and config are committed files. The opt-in relay stores encrypted board records, GitHub id and login, repository names and phone approvals, keeps backups 14 days and deletes boards idle for 90 days.
- Cost at scale
- The software is free; the cost is model usage for every coordinator, builder and verifier you run, since every item is built by one agent and checked by another. No hosted pricing is published.
- Limits
- Not a security boundary between agents on one machine; Windows untested; one month old, 13 open issues and pull requests, and the multi-person shared board is still on the roadmap.
What’s good
- Second-agent verification on the exact commit before anything merges
- Board, spec and rules live in the repo, with no service required
- Published receipts from 908 submissions on its own board
- Lanes own paths, so parallel agents do not collide
- Works with Claude Code, Codex and any MCP client
What’s not
- Created 7 October 2026; still pre-1.0
- Requires adopting SPEC.md, DOCTRINE.md and its workflow
- Hooks skippable with --no-verify; worktree identity is not security
- Windows untested; relay pricing not published
Best for
Developers running several coding agents on a long-lived product with a real test suite, who want agents checked by agents rather than reviewing every diff themselves.
Pullboard alternatives
Pullboard: common questions
- Is Pullboard worth it?
- A serious attempt to make unattended agent work trustworthy, with real evidence from its own board, though it is a month old and asks you to adopt its whole way of working. We rate it 3.7 out of 5.
- How much does Pullboard cost?
- Free under the MIT licence. The optional hosted relay has no published price.
- What are the downsides of Pullboard?
- Created 7 October 2026; still pre-1.0; requires adopting SPEC.md, DOCTRINE.md and its workflow; hooks skippable with --no-verify; worktree identity is not security; windows untested; relay pricing not published.
- Who is Pullboard best for?
- Developers running several coding agents on a long-lived product with a real test suite, who want agents checked by agents rather than reviewing every diff themselves.
- How much does Pullboard cost for a team or in production?
- The software is free; the cost is model usage for every coordinator, builder and verifier you run, since every item is built by one agent and checked by another. No hosted pricing is published.
- Is Pullboard safe for company data?
- Board is a SQLite file in .git; spec, doctrine and config are committed files. The opt-in relay stores encrypted board records, GitHub id and login, repository names and phone approvals, keeps backups 14 days and deletes boards idle for 90 days.
- What are the best Pullboard alternatives?
- The closest alternatives we have reviewed are Vibe Kanban, Kanzo, Spec Kit and Claude Squad.
Answers checked Nov 5, 2026.