Image: GitHub
Why it mattersA team running Claude Code or Codex has nowhere shared to park tasks for the agent. Kanzo gives both sides the same board. The agent picks the next card and reports back in the comments.
A coding agent that writes code has nowhere shared to pick up its next task. Kanzo is a new open-source project that gives the agent a board. The GitHub repository has 133 stars and 9 forks as of 29 October. Version 0.2.1 landed on 27 October 2026 under the AGPL-3.0 licence. The repository was created on 25 October, four days ago.
The project is one program by developer Emanuele Em. It runs a kanban board in the browser at http://localhost:4242. It also runs an MCP server for coding agents at the same address. Both sides see the same cards. The agent moves a card to doing, works on the task, and writes its progress as comments.
What it does
The README names the step for Claude Code: claude mcp add kanzo --scope user -- kanzo mcp. Codex has a one-line equivalent. The agent then takes the next card from the board when the user asks for it. It knows which project to use from the git remote of the repository it is working in.
Keyboard shortcuts match the pattern of other agent-era tools. n adds a task. / or ⌘K searches every project. Tab moves from title to description. ⌘↵ saves and closes.
An optional Backlog column holds ideas the agent should leave alone. A task takes a pasted image, which the agent gets as a file path and can open. New releases install from inside the sidebar.
The numbers in the README
The README lists six measurements, taken on an Apple M4:
- The program is 4.1 MB, with SQLite included.
- Memory sits at 4.5 MB when idle, 8 MB with 3,000 tasks on the board.
- The program starts in 2 ms.
- A board of 3,000 tasks loads in the browser in 17 ms.
- The MCP tool list is about 900 tokens for the agent, sent once per session.
- The agent's own replies are 10 to 60 tokens of plain text.
These are the author's figures. A team on a different machine should run its own numbers.
Sync across machines
The whole board lives in ~/kanzo as kanzo.db and config.toml. The project uses a private git repository the user owns as its sync layer. The user creates an empty private repository on GitHub, or any other git host. They paste its URL into Settings, and the same URL on a second machine. kanzo sync commits the board and pulls or pushes it. If both sides changed, Kanzo asks which to keep.
There is no account, no server, and no third-party service in the sync path. For agents on several machines at once, one board can listen on 0.0.0.0. Every agent then connects to the same http://<host>:4242/mcp endpoint.
What it costs to use it
Installation is one shell line for macOS, Linux and every Raspberry Pi. Every release is signed by GitHub, so gh attestation verify can check a download. The licence is AGPL-3.0, so a company that offers a changed version over a network must share its source.
The project is new. Version 0.2.1 is the current release, from 27 October 2026, two days ago. There are no third-party reviews, no case studies, and no reports of large teams running it in production yet. A team that adopts Kanzo this week is an early user.
A team that already runs Claude Code or Codex may have no shared place for the agent's next task. Kanzo is cheap to try for that team. The usual options are a private chat log and a to-do list the agent cannot see.
Source
- emanuele-em/kanzo, GitHub
- Kanzo v0.2.1 release, GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.