Kaku skill
A skill that teaches an agent inside the Kaku terminal to open panes, type into nearby windows and wait for another agent.
- 6,206 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The agent types commands into other panes and starts processes, so a wrong command in another pane runs as the user.
Install
- Works in
Manual install
curl -fsSL https://raw.githubusercontent.com/tw93/Kaku/main/docs/agent-skill.md -o ~/.claude/skills/kaku/SKILL.mdThe command from the Kaku docs places the skill file for Claude Code. Install the terminal separately: `brew install --cask kaku`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent types commands into other panes and starts processes, so a wrong command in another pane runs as the user.
This is third-party code. Review the repository files before installing.
What it does
Kaku is a fast macOS terminal that shows agent state and notifies you when Claude Code needs you. It ships a `docs/agent-skill.md` file: a skill for an agent running inside Kaku (the `TERM_PROGRAM` variable equals `Kaku` and `WEZTERM_PANE` holds the pane id). The skill describes the `kaku cli` commands. `split-pane` opens a pane with a dev server or tests, `send-text` types into another pane, `get-text` reads its screen, and `list` enumerates panes with their state. `wait` waits until another agent reaches `done`, `blocked` or `error`, and `events` streams state changes. Claude Code 2.1.295 and later and Codex 0.163 and later report what they are doing to the terminal.
Who it is for. macOS users running Claude Code or Codex in the Kaku terminal who want one agent to start servers, tests and a second agent in neighboring panes.
Good fit when
- An agent starts a dev server in a bottom pane and keeps working in its own.
- Tests run in a separate pane, and the result is read from the log after the command exits.
- A task is handed to a second agent for review and the agent waits for its `done` state.
- You need to see which pane has stopped and awaits a user reply.
Not a fit when
- Work happens in iTerm2, Terminal.app or another terminal: the skill is for Kaku only.
- The agent must not touch other panes, while the task needs typing into a window it did not start.
- Linux or Windows is required: Kaku runs on macOS only.
Example request
Run `npm run dev` in a bottom Kaku pane, wait for the server to come up and read the last 50 lines of that pane's screen.Limitations
The skill works only inside Kaku, where `WEZTERM_PANE` is set. A pane closes with its command, so output needed later is saved to a file. Only Claude Code and Codex of the stated versions report state: a plain shell or a build ends a wait only by exiting. The skill's rules forbid closing panes or typing into panes the agent did not start unless the user asks.
How to disable. Delete `~/.claude/skills/kaku/SKILL.md` and start a new session. The Kaku terminal itself stays.
FAQ
How do I install the terminal itself?
Download the DMG from the releases page or run `brew install --cask kaku`. Then open Kaku and check the version with `kaku --version`.
How do I know another agent has finished?
Call `kaku cli wait --pane-id 5 --for done,blocked,error --next --timeout 20m`. The command prints the state that ended the wait and exits 0; exit code 2 means a timeout.
How do I read a pane's output?
`kaku cli get-text --pane-id 3` shows the current screen, and `--start-line -200` adds 200 rows of scrollback.
What is the `--next` flag for?
It counts a state only after the pane has left it. A stale `done` from the agent's previous turn will not end the wait early.
What if a command closes its pane right after exiting?
Redirect output to a file, for example `sh -c 'go test ./... > /tmp/kaku-test.log 2>&1'`, wait with `kaku cli wait --for exit` and read the file with `tail`.
Discussion
No questions or reviews yet.
Related
SkillsThis skill teaches Claude Code and Codex to share the con terminal: inspect panes, run commands in view and coordinate workers
The tty7 skill teaches Claude Code and Codex to drive tty7 terminal panes: run commands in a real PTY, read the screen and hand work to another agent
Over twenty skills for coding agents: MCP and hooks, plan mode, a bug-fix protocol, prompting and a bash safety hook
A macOS terminal with the termio CLI and a skill: agents spawn sibling sessions and you see which one needs your answer
A mirror of Lauren Tan's pstack set from Cursor: over 50 skills, the poteto-mode with 23 playbooks and 24 principles for strict agent work on code
Softaworks Agent Toolkit
Softaworks Agent Skills
A marketplace of skills, subagents and commands for coding agents: development, documentation, planning and workplace communication
README in short
Kaku is built in Rust and described as a fast terminal with zsh completion, syntax highlighting and agent state shown at the top right. The skill sits in the docs next to the `kaku cli` reference.