Pullboard CLI
A CLI for AI agents sharing a repository: a task board, a spec and acceptance by a second agent against a specific commit
- 721 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The tool installs git hooks, refuses commits outside a lane and runs your check command before every submission and push
- The `pullboard run` command lets a lighter model build simple items unattended
- Phone pairing goes through a third-party relay server, so make sure the board holds no secrets
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/pullboardDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The tool installs git hooks, refuses commits outside a lane and runs your check command before every submission and push
- The `pullboard run` command lets a lighter model build simple items unattended
- Phone pairing goes through a third-party relay server, so make sure the board holds no secrets
This is third-party code. Review the repository files before installing.
What it does
Pullboard installs as an npm package and sets up a working order for several agents in a git repository. The `pullboard init` command creates a `SPEC.md` file with requirements, a `DOCTRINE.md` file with development rules, agent instructions, git hooks and a task board. The main board is a SQLite file inside `.git` and does not enter commits. A separate relay supports phone and other-machine access. Each requirement in `SPEC.md` is one line with an identifier, a status and a check, for example `G1 [approved, must] ... | gate: test/checkout.test.js`. Only rows you approved count. Work is filed as an item that a builder agent claims, and the acceptance bar is frozen at the moment of the claim. Each agent works in its own lane, a separate git worktree, and a commit outside the lane is refused. Work can be submitted only when the check from `pullboard.json` passes, for example `npm test` or `pytest`. A second agent then verifies that exact commit and either accepts it with evidence or sends it back with a reason. Questions travel up the chain to a coordinator, and only decisions that belong to a human reach you. For Claude Code the repository ships skills, for Codex and other agents there is `AGENTS.md` and the `pullboard prompt <role>` command. `pullboard view` shows the state of every project on the machine.
Who it is for. For developers and founders who run a long project with several Claude Code or Codex agents and want human decisions to survive between sessions.
Good fit when
- A project runs for months and in every new session agents forget your decisions about what counts as done
- Several agents edit one repository and need separate lanes so they do not overwrite each other
- You want a second agent to verify an agent's "done" on the exact commit without you reading every line
- You need tests to be a mandatory gate for submission and a push to be refused when the check is red
- You want a board that shows what is being built, what is in review and which decisions wait for you
Not a fit when
- You work with a single agent on a small edit: a spec, lanes and second-agent review add steps you do not need
- You use Windows: project CI runs on Linux and macOS and Windows is not tested yet
- You need protection from a hostile agent: worktree identity holds on one machine and is not a security boundary
- You lack Node.js 22.13 or newer and git: the package cannot be installed without them
Example request
Install Pullboard and set it up here: the repository is an online shop, let the gate be `npm test` and you act as the coordinatorLimitations
Git and Node.js 22.13 or later are required. Initialization changes project instructions and hooks, so preserve existing settings first. Worktrees separate tasks but do not isolate agents from files; local Git hooks can be bypassed. Evaluate the actual checks and the second reviewer’s findings.
Enabling the relay sends encrypted board snapshots to an external server. Its operator sees the GitHub account, repository names, devices and IP addresses. Phone-approval context, including exact command text, is stored in plaintext. Keep secrets out of it. The author’s encryption design does not make all metadata private.
How to disable. Remove the installed hooks from `.git/hooks`, uninstall the package with `npm uninstall -g pullboard`, then delete `SPEC.md`, `DOCTRINE.md` and `pullboard.json` from the repository if you no longer need them. `pullboard export` makes a backup of the board first.
FAQ
How do I install Pullboard and try it on a test repository?
Run `npm i -g pullboard`, then `pullboard tour`. The command plays a full cycle on a throwaway repository: you approve one rule, a builder builds it, a verifier rejects the first version, the builder adds a test and the coordinator merges the work with a receipt.
How does the spec differ from the doctrine?
The spec says what to build: "a shopper can pay by card in one step". The doctrine says how to build it: "we keep only the data we need". The spec lives in `SPEC.md` and the development rules in `DOCTRINE.md`.
What are lanes and how does an agent claim work?
A lane gives an agent paths and spec sections in `pullboard.json`. `pullboard worktree <lane>` creates a worktree and prints first instructions, and `pullboard next` claims the next free item according to the board’s current state.
Where is the board data stored?
The spec, doctrine and config are plain files in the repository. The board itself is a single SQLite file inside `.git`, shared by every worktree and never committed. Submitted work is pinned in git, so it survives a deleted worktree.
Does Pullboard work with Codex?
Yes. Claude Code gets skills from the `skills/` folder, while Codex and other agents read `AGENTS.md` or the text from `pullboard prompt <role>`. Any MCP client can be added too.
Discussion
No questions or reviews yet.
Related
CLI toolsOne config for Claude Code and Codex in every repo: shared rules, skill packs, second-model review and a destructive-command guard
A memory and coordination layer for coding agents: specs and issues live in git while agents work in parallel
A bridge between local CLI agents: Codex, Cursor or Claude Code delegate work to Antigravity, Grok, Kimi and seven more.
A Rust terminal task board: every task runs in its own git worktree and tmux window while Claude Code, Codex, Gemini CLI and other agents work in parallel
DevChain runs a team of AI coding agents on your own machine with a shared board, chat, code review and live terminals.
Shared team memory for people and AI agents in Git: decisions and handoffs live in Markdown, reachable from Claude Code and Codex.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Pullboard (this card) | no | 721 | |
| oh-my-claudecode | no | 40,625 | multi-agent orchestration inside Claude Code with roles and an autopilot |
| comanda | no | 332 | an orchestrator driven by YAML scenarios that repeat until the result is accepted |
README in short
The project uses MIT. The tour command demonstrates a task moving between builder and verifier in a separate repository before you configure your own project.