Clawd on Desk: a pixel crab that turns coding agent hooks into animations and approval cards

Twelve animated states fed by agent hooks and logs

Clawd on Desk is an Electron app whose pet changes state according to what your coding agent is doing. The source of truth is the agent itself: hooks and logs drive 12 animated states. Clawd thinks while you write a prompt, types while tools run, juggles once subagents pile up, pops a card when a tool needs approval, and dozes off when you step away. On top of that sit eye tracking, a sleep sequence, reactions to clicks, and a mini mode that peeks in from the screen edge. Which event triggers which animation is set out in the state mapping guide under docs/guides. The point of all this is narrow and practical: start a long task, walk away, and let the crab tell you when it is done or blocked.

Claude Code and Codex are hooked on first launch, the rest by click

The supported list covers more than 20 agents, including Claude Code, Codex CLI, Copilot CLI, Gemini CLI, Cursor Agent, Kiro CLI, Qwen Code, opencode, Kimi Code CLI, Qoder, DeepSeek Harness and MiniMax Code. Each session is tracked side by side, so two agents in two terminals show up separately.

Only Claude Code and Codex CLI are connected automatically on first launch. Every other integration is installed from Settings, then AI Coding Tools, after which Clawd keeps it in sync. A tool that has no integration yet can still report state through the custom HTTP agent endpoint, documented in docs/guides/custom-agent-http.md. The consequence of the automatic step is worth knowing before you launch: running Clawd changes the hook setup of Claude Code and Codex on that machine, whether or not you planned to watch both.

Permission bubbles do not exist for every agent

The feature that earns Clawd a place on screen is the permission bubble. When an agent asks to run a tool, a floating card offers Allow or Deny, and a global hotkey does the same without reaching for the mouse. The permission handling section of the setup guide offers three modes: ask every time, question prompts only, or auto-approve.

This does not work the same way for every supported agent. Gaps such as agents without permission bubbles are collected in docs/guides/known-limitations.md, and for those agents Clawd can show status but cannot gate tool calls. If approval from the pet is the reason you want it, check your agent against that file first. Auto-approve, for its part, removes the human check entirely, so it hands every tool request straight through.

The phone view watches sessions but cannot approve them

Away from the desk there are three channels, and they do different things. Telegram and Feishu or Lark can approve requests remotely. Slack receives notifications. The mobile PWA, built from the pwa/ directory in the repository, lets you follow sessions on your phone but is read-only. A blocked agent waiting for permission cannot be unblocked from the PWA; you need Telegram, Feishu or Lark set up for that.

Back at the desk, a HUD sits next to the pet and a dashboard lists every live session with a one-click jump to its terminal, quota rings and a private local recap. Clawd also follows agents beyond the local machine: Remote SSH delivers state and permission bubbles from agents on a remote server, and a separate guide spells out what is supported for Codex inside WSL.

Installers per platform, and where Homebrew stops

Prebuilt installers come from GitHub Releases. Windows gets Clawd-on-Desk-Setup-<version>-x64.exe or the -arm64.exe variant. macOS gets a .dmg, or a cask:

brew install --cask clawd-on-desk

Linux has two formats with different update paths. The .AppImage has in-app updates. The .deb is installed through an APT setup described in docs/guides/apt-setup.md, which also covers migration. Homebrew works on Linux too, but only on x86_64. The package.json build scripts mirror these targets: nsis:x64 and nsis:arm64 for Windows, dmg and zip for each Mac architecture, and AppImage:x64 deb:x64 for Linux. Note that a bare npm run build runs electron-builder --win and produces only Windows output. Releases have come quickly: v1.3.0 on 2026-10-10, v1.4.0 on 2026-10-19 and v1.5.0 on 2026-10-31, with the last push on 2026-11-03.

Running from source pulls Electron and registers hooks

A source install is meant for contributing, testing unreleased code or debugging integrations. It downloads Electron and the build tooling, so node_modules gets large:

# Clone the repo
git clone https://github.com/rullerzhou-afk/clawd-on-desk.git
cd clawd-on-desk

# Install dependencies
npm install

# Start Clawd (auto-registers Claude Code and Codex hooks on launch)
npm start

npm start runs node launch.js, while the Electron entry point is src/main.js. A postinstall step runs scripts/verify-electron-install.js, so a broken Electron download surfaces during npm install rather than at launch. The repository pins its Node version in .nvmrc, and prerequisites plus macOS and Linux notes are in docs/guides/setup-guide.md. The package.json description still reads as a pet that reacts to Claude Code sessions, a leftover from before the agent list grew.

Crab, calico cat, cloud, and themes you draw yourself

Three characters ship in the box: Clawd the pixel crab, Calico the three-colour cat and Cloudling. Hash Sage and Whale-chan are optional official themes downloaded from Settings, then Theme, which is also where a Codex Pet zip can be imported. Each theme carries the same core states, from idle and thinking to typing, building, juggling and conducting.

To make your own, scaffold it with:

npm run create-theme -- my-theme

and follow docs/guides/guide-theme-creation.md, which covers building and validating it. A community tool, Live2Pet by receyuki, converts Live2D models into Clawd themes. The interface is translated into 7 languages, and the README exists in English, Simplified and Traditional Chinese, Korean, Japanese and Spanish. The code is AGPL-3.0, with a NOTICE.md alongside the LICENSE file.

Editorial conclusion

Clawd on Desk suits developers who leave Claude Code or Codex running long tasks and want a visible signal plus a fast Allow or Deny for tool requests, on Windows, macOS or Linux. It suits less well anyone who needs approvals from their phone, since the PWA is read-only, or who relies on an agent listed in the known limitations guide as lacking permission bubbles. Before installing, check that guide for your agent, and know that the first launch registers Claude Code and Codex hooks.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. Release notes