Every agent.
One screen.
Leave them running.

A tmux for coding agents. KanBot owns a real terminal for every agent you run — Claude Code, Codex, Gemini, anything with a CLI — and shows them all in one terminal app or one web board. See who is working and who is stuck. Type into any of them, from a terminal or your phone. Hand them a queue and walk away.

$pipx install kanbot && kanbot

zero-install: uvx kanbot · MIT · Python 3.9+ · macOS & Linux

—GitHub stars —Forks —Downloads / month —Last commit
kanbot — laptop · gpu-box

    

  


  
Claude CodeCodexGemini CLIOpenCodeAiderCursor AgentGLM / Z.aiKimiHermesany shell


  

    
How it works

    
The runner owns the terminals. You just visit.

    
A background runner spawns each agent inside a PTY it controls. Close the tab, sleep the laptop, ssh in from somewhere else — the process is still there, the scrollback replays, and you pick up exactly where it was.

    

      
01
Panes that outlive you
Each agent gets a real pseudo-terminal held by the runner, not by your browser tab. Attach from the web board, from your phone, or with kanbot attach. Ctrl-] detaches; the agent keeps going.

      
02
Exact state, not spinners
Claude Code reports working / blocked / idle through its own lifecycle hooks — no guessing. Every other CLI is classified from what is on its screen.

      
03
Blocked comes to you
An agent waiting on a permission prompt jumps to the top of the board, flashes the tab, and pings you: a macOS banner by default, or any command you configure — Telegram, ntfy, Slack.

      
04
Two front doors
kanbot opens a terminal app: agents on the left, the live agent on the right, Enter to type into it, Ctrl-] back. kanbot up opens the web board with one-tap y n esc keys — for answering "Allow this command?" from a phone at dinner.

      
05
An API your agents can drive
The CLI and the local socket are one surface: agent.start, agent.wait --until idle, agent.read, agent.prompt. Scripts and other agents orchestrate agents without faking keystrokes.

      
06
Still a queue
Cards, multi-step workflows, review gates, and a 10-hour goal spree with a real shell stop condition. All of it runs inside panes now, so unattended runs are watchable and interruptible live.

    

    

      
working
Tokens flowing, tools running. Leave it alone.

      
blocked
Waiting on a human: a permission, a question, a trust prompt.

      
idle
Finished the turn, ready for the next prompt.

      
done
Process exited. Exit code and scrollback kept.

    

  


  

    
Compare

    
tmux keeps processes alive. KanBot knows what they're doing.

    
Terminal multiplexers solve persistence. Agent multiplexers add state. KanBot adds the part that lets you leave: a queue, a phone-friendly board, and a way to be told when an agent needs you.

    

      
tmuxHerdrKanBot

      
        
Processes survive disconnectyesyesyes

        
Agent state (working / blocked / idle)—yesyes, hook-exact for Claude Code

        
Terminal app (sidebar + live agent)panes, no stateyesyes

        
Web board, works on a phone——yes

        
Notified when an agent is blocked—in-appbanner, or any command (Telegram, ntfy…)

        
Socket API + CLI for scripts and agentssend-keysyesyes

        
Task queue, workflows, gates, long autonomous runs——yes

        
Resume yesterday's session from disk—yesyes

        
Split panes and layouts in one terminalyesyesnot yet — the board is the layout

        
Installbrew / aptcurl | sh (Rust)pipx install kanbot (Python)

      
    

    
Honest gaps: KanBot has no native TUI and no pane splitting. If you live in one terminal window, Herdr is excellent. If you leave agents running and check on them from elsewhere, KanBot is built for that.

  


  

    
Quickstart

    
Three commands. Then go do something else.

    

      
1
Install and open the terminal app. It starts the runner in the background on first run.
pipx install kanbot
kanbot

      
2
Give an agent a task — from the board (press n) or the CLI. It opens as a live pane.
kanbot agent start claude \
  "fix the flaky test" --cwd ~/repo

      
3
Leave. Come back from any terminal or any browser; answer a prompt; leave again.
kanbot ps
kanbot attach 3f2a
kanbot agent keys 3f2a y Enter

    

  


  

    
CLI

    
Herdr-shaped commands, so the muscle memory transfers.

    

      
kanbotthe terminal app · Enter focus · Ctrl-] back · n new · q quit

      
kanbot upserver + local runner in the foreground, opens the web board

      
kanbot runner --server http://host:8787add another machine

      
kanbot psevery agent here: state · agent · cwd · title

      
kanbot attach <id>your terminal becomes that pane · Ctrl-] detaches

      
kanbot agent start claude "…" --cwd ~/repoopen a live TUI (--headless for print mode)

      
kanbot agent start claude --resume <session>pick up a session from disk

      
kanbot agent wait <id> --until idleblock until it's ready · blocked · done

      
kanbot agent read <id> --lines 40the screen as plain text

      
kanbot agent prompt <id> "now add tests"type + Enter

      
kanbot agent keys <id> y Entertmux-style key names · C-c · Escape · Down

      
kanbot agent kill <id>HUP, TERM, then KILL

      
kanbot review --gatemulti-agent code review as a workflow gate

    

    
Same methods over the unix socket at ~/.kanbot/runner.sock and over HTTP from the board: GET /api/panes · POST /api/panes/start · POST /api/panes/{id}/input · GET /api/panes/{id}/read.