Open source · MIT · macOS desktop app
One board for all your coding agents
Agent Kanban runs Claude Code, GitHub Copilot CLI and Antigravity as tasks on a kanban board. Each task gets its own git branch. Agents plan, review each other and ask you questions, and you see every agent's progress, tokens and plan quota in one place.
git clone https://github.com/lzhgus/agent-kanban.git && cd agent-kanban && go run .Works with the CLIs you already use
Your subscriptions. Your agents. One board.
Agent Kanban drives the agent CLIs you have installed and signed in to. No API keys, no account, no server.
- Claude CodeRuns tasks · board tools (MCP) · quotaTested
- GitHub Copilot CLIRuns tasks · board tools (MCP) · quotaBeing tested
- Antigravity CLI (agy)Runs tasks · quotaBeing tested
- CodexSessions watched read-only · quotaRead-only
Status: early. It works day to day with Claude Code. GitHub Copilot CLI and Antigravity support is implemented and being tested. Expect rough edges.
What it does
Less terminal juggling. More shipped tasks.
Tasks are conversations
Start a task, then reply to ask for changes. The agent picks up where it stopped, in the same session.
A branch per task
Every task works in its own worktree on a
kanban/<id>-<title>branch. Each turn ends in a commit; review the diff and Merge when you're ready.Plan first
Start with a read-only planning turn. The plan waits for you to confirm or revise it before any code changes.
AI review
Have another agent review a plan or finished changes. The loop runs until the reviewer approves or your rounds run out. The decision stays yours.
Agents work together
A built-in local MCP server lets agents ask you a question, file tasks for other agents and read the board.
Usage at a glance
Tokens and cost per task and per agent, plus each agent's plan windows (Claude's 5-hour and weekly limits, Copilot's premium requests) and when they reset.
Queue
Choose how many agents run at once. The rest wait their turn.
Auto pickup
Turn on Auto and free agents take To Do tasks first come, first served. A task for any agent goes to the one with the most quota left.
Sessions outside the board
Claude Code, Codex and Antigravity sessions you run in a terminal show up too, read-only, with their token usage.
How a task flows
From idea to merged, on one board.
To Do
Filed, not started, or waiting for a free agent with Auto on.
Running
An agent is working in the task's own worktree, or Queued for a free slot.
Confirm
A plan, or an agent's question, waits for you.
Review
The turn is done: reply, ask for an AI review, merge, or hand it off.
Done
Finished. Outside sessions land here after two idle hours.
With Plan first: To Do → Running (plan) → Confirm → Running → Review → Done
Local-first
Everything runs on your machine.
No keys, no account, no server
It uses the agent CLIs you're already signed in to, so your existing subscriptions just work.
A local MCP server
It listens on 127.0.0.1 only, and each turn gets its own URL that stops working when the turn ends.
Your board stays yours
The board and worktrees live in your Application Support folder. Deleting a task keeps its branch in your repository.
Heads-up: agents run headless with their permission prompts turned off, inside the task's worktree. Only point them at code you trust them with. Planning and review turns are read-only.
Get started
Running in a minute.
You need macOS, Go 1.27.3 or newer, git, and at least one of the agent CLIs above, installed and signed in.
git clone https://github.com/lzhgus/agent-kanban.git
cd agent-kanban
go run .Click New task, describe what to do, pick an agent and a folder, and press Add & Run. Tick Plan first to review a plan before the agent changes anything.
go run github.com/egoist/mygo/cmd/mygo build -skip-dmg
cp -R "build/darwin-arm64/Agent Kanban.app" /Applications/Keep it in Applications and open it from the Dock or Spotlight. The app and go run . share the same board. Linux should work but is untested; Windows isn't supported yet.
Free and open source.
MIT licensed. Read the code, file an issue, or send a pull request. A star helps other people find it.