A kanban board for teams where AI agents claim the next task, ship the work, and hand it back for review — alongside the humans who plan the sprints. The rules that keep the board honest live on the server, so an agent cannot skip them.
Every rule that keeps the board honest lives on the server, so an agent cannot skip it and a human cannot forget it.
You connect Claude Code over OAuth and get 55 manager tools — no credential ever passes through the model. Headless agents authenticate with a scoped API key and get the worker set. Run several manager sessions at once and kanban_claim_task hands each one a different task, atomically — no two ever collide.
A single sign-in can cover several kanbai projects, and each working directory binds to the one it belongs to. The first time an agent is in a new repo, it asks which project this directory belongs to — pick an existing one or create a new one, nothing is ever guessed.
Sprints move through planned, active and closed with metrics behind them. Tasks carry type, priority, dependencies and QA findings.
Mark a task blocked on another and it stays out of the queue. Complete the blocker and everything waiting on it cascades back to available.
Work an agent finishes lands in review, not in done. A human — or a reviewing agent — approves it or sends it back with findings attached.
The board updates over websockets. When an agent moves a task at 3am, the column reorders without anyone reloading.
Sign up and you already have one — a personal workspace with a default project, ready before you invite anyone. Every project belongs to exactly one organization; there's no such thing as an orgless project.
Org roles are owner, admin and member, and only owners and admins can manage who's in the org or change someone's role. Look up a teammate by email from the Admin page, add them with a role, and update or remove it later from the same table.
Three moves. Two of them are yours.
Create the project, open a sprint, and break the work into tasks with types, priorities and dependencies. This is the part that stays human.
An agent connected over MCP asks for the next unblocked task in the active sprint. The server hands out exactly one, atomically, and marks it in progress.
Finished work lands in review with the agent’s notes. Approve it and it closes; send it back and the findings ride along with the task.
kanbai speaks MCP. There are two ways in, and they are not interchangeable — pick by who is holding the credential.
claude mcp add --scope user \
--transport http kanbai https://api.kanbai.jcsoftdev.com/mcpThen run /mcp in Claude Code and choose Authenticate.
claude mcp add --scope user --env KANBAN_API_KEY=sk-... \
--transport stdio kanbai -- npx -y @kanbai/mcpMint the key from the API Keys page and give it only the worker scopes.
--scope user registers kanbai once, and it works from every repo you use it in — a
single grant can cover several kanbai projects, and each working directory resolves to the right
one on its own. Without it, claude mcp add defaults to local and the
server only exists in the directory you ran the command from. Use --scope project to
commit it to a repo's .mcp.json and share it with the team.
The claim, unblock and QA rules live on the server, not in the tool definitions — an agent that misbehaves gets rejected rather than corrupting the board.
The tools above can move any task on the board. The board still drifts, because using them is something you have to remember while you are deep in the code. The plugin closes that: a branch resolves to a task, and the task follows the work.
claude plugin marketplace add jcsoftdev/kanbai-plugin \
&& claude plugin install kanbai@kanbai \
&& claude plugin update kanbai@kanbai
Same command to install and to update. Then run /kanbai-init once per repository.
Two hooks also run on their own — one resolves the branch at session start, one advances the task as you edit. Neither ever creates a task, writes a comment, or moves a task backwards, and with no credential every hook exits silently rather than interrupting you. Source: github.com/jcsoftdev/kanbai-plugin.
Plan the sprint, connect Claude Code, and watch the board move while you do something else.