作者打造 Kanbell，集中追蹤多個 Claude Code 智能體

A few months ago my typical afternoon looked like this: four terminal tabs, four Claude Code sessions, four tasks. One was writing tests. One was refactoring a route. One had finished twenty minutes earlier. And one had been sitting on a question ("Use zod or valibot?") for half an hour while I stared at a different tab.

Running agents in parallel is easy. Knowing which one needs you is the hard part.

So I built Kanbell, a Mac app for people who run several coding agents at once.

What it does

Start from a task, not a blank prompt. Kanbell has a board. You pick a task, and Kanbell starts the agent in its own terminal and pastes the brief for you:

Brief prepared by Kanbell from a board issue:
#128 — Rate-limit uploads

Limit POST /uploads to 20 requests per minute per API key.
Return 429 with a Retry-After header.

One terminal per agent, and the one waiting on you lights up. Every session gets a real terminal, so you can still type in it like you always do. When an agent stops to ask you something, its session lights up in the app and in the menu bar. You don't have to poll your tabs any more.

A live map of every agent and subagent. Claude Code spawns subagents. Kanbell shows them as a tree, live, so you can see what is actually running and on which task.

Answer from the overview. The overview window shows all your agents as cards. Double-click one and the card becomes its real terminal, in place. You answer the question right there, without hunting for the right tab.

Review the diff next to the terminal that wrote it. A Git panel sits beside each session: files changed, additions, removals. You review the work where it happened.

Which agents?

• Claude Code: full tracking: working, waiting, subagents, usage.
• Codex CLI: the same, without the subagent tree.
• Gemini CLI: working, waiting and permission requests. It exposes no quota or subagents.
• Any other CLI agent: runs in the terminals, untracked.

Kanbell runs your agent, it doesn't replace it. You keep your own subscription, and usage stays on your account.

What it doesn't do

• It doesn't send your code anywhere. There is no Kanbell account and no Kanbell server. The app only talks to Polar (license) and GitHub (updates).
• It doesn't run agents for you in the cloud. Everything runs locally, in your terminals, with your CLI.
• It doesn't hide the terminal. If you like typing to Claude directly, you still can. Kanbell adds the board, the tracking and the review around it.

How it's built

Kanbell is an Electron app with React, Tailwind and xterm.js for the terminals. Each session runs in a real pseudo-terminal (node-pty), so the agents behave exactly as they do in iTerm or Terminal.app.

The part I'm proudest of is the tracking. Kanbell doesn't guess an agent's state by scraping the screen. It reads the signals the CLIs themselves expose (for Claude Code, its hooks), so "waiting for you" means the agent really stopped and asked.

And yes: most of Kanbell was built with Kanbell. A board of tasks, a few agents in parallel, and me answering whichever one lit up.

Try it

• Requirements: an Apple Silicon Mac, git 2.45 or newer. Full tracking needs Claude Code 2.1.80 or newer.
• Trial: 7 days free, with no card and no account. The trial runs inside the app.
• Price after that: 12 CHF a month or 96 CHF a year (about $15 a month), for every feature, all updates and up to 3 Macs.
• Linux and Windows: a Linux beta is starting with a few testers, and Windows is in the works.

👉 kanbell.com

I'd love feedback, especially from people who already juggle several agents a day. What is your current setup: tmux, worktrees, a wall of tabs? And what's the moment you lose track? Tell me in the comments, or come say hi on Discord.