When people see "Kanban," they usually think of another annoying tool humans have to manually update.

This isn't that. You don't manage this board your AI agent does.

I built this because I was hitting a wall with Claude Code and OpenCode.

Once you get 40 prompts deep into a massive feature, the context window gets bloated ( context rot ), the agent gets sluggish, and every prompt starts draining your token budget.

You want to run /clear to get a fresh, fast context window, but if you do, the agent gets instant amnesia.

So I built Kanban, which is a local-first MCP server.

It gives the AI an external working memory layer that lives outside the chat log.

Instead of writing to a messy text file, the AI uses MCP tools to programmatically organize its own thoughts into a structured Kanban schema.

This lets you do something awesome:

You can type /clear right in the middle of a complex refactor to drop your token usage back to zero. Then, your very next prompt to the blank-slate agent is just:

"Check your board and resume."

The agent calls task_recall , pulls the exact card it was working on, reads the timestamped notes it left itself before the context wipe, and picks up right where it left off.

Max reasoning speed, zero token bloat.

I only built the web UI (running on localhost) so I could visually spy on what the AI is thinking.

You just sit back and watch the agent create cards, append notes, and drag tasks from "In Progress" to "Done" in real-time as it runs terminal commands.

If it gets stuck, it moves the card to "Blocked," writes a markdown note explaining why, and hands the terminal back to you.

It’s 100% local, stores everything in plain JSON, and automatically swaps project workspaces based on whatever Git repo you cd into.

It's basically a workflow tool for robots, but we get a cool dashboard to watch them work.

Here's the code if you want to check it out:

{

"mcpServers": {

"kanban": {

"command": "npx",

"args": ["@cwim/kanban", "mcp"]

}

}

}

https://github.com/cwimhq/kanban

[留言]

为什么值得关注

能改变理解方式，而不只是重复常识；符合当前抓取需求；它提供了新的理解或解释，而不只是表面观点

来源：reddit，领域：projects，保留分：0.66