v0.3.0 on npm
kabane
A fast issue tracker you never have to use. Unless you want to.
Your coding agents claim, update and close the issues over MCP while they work. When you want to look, the board in your terminal shows what’s open at a glance. It all lives in one SQLite file on your machine.
bun add -g kabaneAlex asks Claude Code for the next kabane issue. It finds JMYA-2 and reads the description and Alex’s note in one kabane_context call.
What you get
Your agents do the paperwork
Claude Code, Codex and Gemini CLI list, claim, comment on and close issues over MCP. One
kabane_contextcall hands an agent the whole brief.See who did what
Each move, comment and work log names its author: you, or the agent that made it.
Press A and an agent takes it
From the board,
Ahands an agent the issue, the rows you marked and your filters./triage,/splitand/duplicatesare a slash away.Hand off a whole planNew in 0.3
kabane loopworks through a parent issue’s subtasks in dependency order, across agent sessions. Your checks run after each one, and any question waits for you on the board.No round trip to a server
Issues live in one SQLite file on your machine, so reads and writes are local. No account to make, and one machine is a complete setup.
Your sessions learn to dispatch
/setup-dispatchreads your project’s CLAUDE.md, AGENTS.md and agents, then writes a dispatch skill, so your coding sessions turn asks into issues and hand them out.
From a quick fix to a whole wave
A quick fix
Say what you want. The session files one issue, does the work and closes it.
A bigger change
Dispatch splits it into subtasks and sends them out in waves, in dependency order.
The wrap-up
Review findings and the docs work land as issues too, so nothing falls off the board.
How it’s built
One TypeScript core runs on Bun on your machine and on Cloudflare Workers in the hub.
- Board
- OpenTUI, made for a 40-column pane
- Storage
- SQLite on each machine, an append-only log between them
- Agents
- MCP over stdio, or over HTTP through the hub
- Copilot and loop
- Claude Code, Codex or Gemini over ACP
- Hub
- A Worker and a Durable Object, behind Cloudflare Access
Add machines with a hub you host
- Each machineThe database, the board and the MCP server. Enough on its own.
- Your hubA Cloudflare Worker, behind Access, that relays changes and serves MCP.
- Everything elseYour other machines sync through it. ChatGPT and Claude.ai connect to it.
Questions
- Is it free?
- Yes. Kabane is MIT licensed, and your agents use the plans you already pay for.
- Do I have to use the board?
- No. Your agents can do all of it over MCP. The board is for when you want to see where things stand: keyboard first, made for a narrow terminal pane.
- Do I need the hub?
- No. One machine on local SQLite is a complete setup. The hub syncs several machines and lets ChatGPT and Claude.ai in, and you deploy it to your own Cloudflare account.
- Does it replace Linear or GitHub issues?
- No. It’s for one person and their agents. Team work stays in Linear or GitHub, and Kabane links to it.
- Which agents work with it?
- Claude Code, Codex and Gemini CLI on your machine, wired up by setup. Through the hub, Hermes, ChatGPT and Claude.ai too.
- Where’s my data?
- In a SQLite file under
~/.kabaneon your machine. With a hub, changes also go to the Worker you deploy. - Is it safe to let a loop run?
- Your checks run after each subtask and reopen it when they fail. A worktree keeps checkpoints you can roll back to, and the loop never answers a permission request for you: it waits on the board.
Your first issue, closed by an agent
- Install Bun 1.4 or later, then run
bun add -g kabane. - Run
kabanein a project. Setup asks your name and connects the agents it finds. - Open an agent session and say “take the next kabane issue”.
Next
- Extensions from anyone: put a
kabane-<name>on your PATH and it becomes a command. - One loop across harnesses, Claude Code on one subtask and Codex on the next.
- One feed for every running loop.
- A board that reads as well on a light terminal as on a dark one.
Part of what I’m building
- 3pewTracked in KabaneMy family’s private dashboard. Each new screen is a short JSON definition that an agent writes and a person publishes.
- tokebakTracked in KabaneA Rust harness for voice and text agents: a speech model holds the conversation, a backend agent runs the lookups and cites its sources, and longer jobs go to other agents.
- icitteTracked in Kabane, and shows Kabane issues as cardsA workspace you can talk to: what you say becomes notes and cards on a canvas, coding agents take the longer work, and the notes and boards are plain Markdown and JSON Canvas files.
- scopecraft-commandThe first try at agents working a task listA markdown task manager with an orchestration harness on top: parallel Claude Code sessions, one git worktree each, dispatched through tmux and watched from a dashboard.