All Pet — AI skill for Claude Code

AllPet 🐾

**One desktop pet that watches all your AI coding agents — Codex, Claude Code / Desktop, DeepSeek Harness (DSH), and Grok — with a cross-platform status / watch CLI.**

The native pet GUI is macOS; the cross-platform Electron shell (`desktop/`) shows the pet plus a graphical pet manager on macOS, Linux, and Windows. Core monitoring and the `status` / `watch` CLI build and run everywhere.

The pet switches animation based on what your agents are doing — idle, running, waiting, done, failed — and each task bubble lets you jump straight back into the originating session.

[中文文档](./README.zh-CN.md) · English

---

Why one pet for all these platforms?

What you actually use every day is not "a model" — it's a **harness × model** combination. The harness decides what context the model sees, which tools it gets, and when it retries or wraps up. The same model behind a different harness can produce very different results:

• The same Claude Sonnet 4.6 scores ~71% on SWE-bench Verified inside Claude Code, but only ~52% behind the Continue shell. (TensorFeed)
• The same Claude Opus 4.5 scores 45.9% under the unified SEAL scaffold, and 55.4% back in its own Claude Code. (arXiv 2605.23950)
• The same Grok 4 scores 58.6% under a generic SWE-agent, and 72–75% with xAI's own scaffold. (arXiv 2605.23950)

The pattern is consistent: **each model is strongest inside its own harness.** So "Claude in Claude Code, GPT in Codex, Grok in Grok" is the everyday reality — which is exactly why AllPet watches Codex, Claude Code, DSH, and Grok at the same time.

Features

• One pet, four platforms — natively reads Codex, Claude Code / Desktop, DSH, and Grok task state from your local session logs (no accounts, no API keys).
• Live state animation — the pet switches between idle / running / waiting / done / failed as tasks progress.
• **Task bubble