SkillFoxx

GPUtw Skill

The official GPUtw skill and MCP server: an agent rents a GPU in Taiwan, runs a command and shuts the instance down

  • 119 ★ on the repository
  • Last commit
  • MIT
SkillOfficial

Broad access

Why this level

  • The agent creates and stops paid GPU instances and spends the account balance
  • The API key opens instances, storage and commands inside the machine
All reasons and checks

gputw-ai/gputw-skillAuthor:GPUtw.ai

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add skills/gputw-skill

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Broad accessThe tool can change data, run commands or handle secrets.

  • The agent creates and stops paid GPU instances and spends the account balance
  • The API key opens instances, storage and commands inside the machine

This is third-party code. Review the repository files before installing.

What it does

GPUtw Skill teaches an agent to work with the REST API of the GPUtw.ai cloud, where GPUs are rented. The user writes in plain language, such as "open an RTX 4090 with the PyTorch template, print the SSH command and stop it at the end". The agent picks a card and node, creates an instance, polls `/status` until `RUNNING`, runs the command, moves models into persistent `/vault` storage and stops the machine. The kit holds a `SKILL.md` with a decision tree, twelve guides on auth, instances, ports, vault uploads and errors, an endpoint reference, a Python client and the `/gputw-deploy`, `/gputw-monitor`, `/gputw-vault` and `/gputw-debug` commands. It also bundles the `@gputw/mcp-server` MCP server with 22 tools. The agent cites only endpoints from the list in the repository and calls anything else undocumented. The knowledge base is written in Traditional Chinese, and answers come in the user's language.

Who it is for. Developers and researchers who run training and inference on cloud GPUs and want an agent to handle renting, starting and stopping.

Good fit when

  • You want to bring up an RTX 4090 in one go, wait for `RUNNING` and get the SSH command
  • A Hugging Face model must be downloaded straight into storage where ComfyUI can see it
  • A machine is stuck in `DEPLOYING` and you need to read its status, events and logs
  • CI needs a separate key that may only upload files to the vault

Not a fit when

  • You need another cloud provider: the skill knows only the GPUtw API
  • You need actions available only in the GPUtw console: the agent cannot perform them
  • You have no `gputw_live_` key: protected calls are unavailable and only the catalog and templates are open

Example request

Open an RTX 4090 on GPUtw with the PyTorch template, print the SSH command once it is RUNNING, and always stop it at the end.

Limitations

Paid operations need a `gputw_live_` key with the right scopes. Deploying works with the deploy preset plus `instances:manage`, and `instances:exec` is best given to a dedicated key. Rentals spend account balance: an hour-ahead balance rule applies before a machine starts, and the API answers `402` when funds are short. The guide tables are a snapshot from September 2026, and the agent reads the current spec from the documentation site. Some actions are browser-only and the skill does not try to work around them.

How to disable. Stop any running instances, remove the gputw plugin through /plugin or the ~/.claude/skills/gputw folder, and revoke the key in the GPUtw console. Remove the MCP server with claude mcp remove gputw.

FAQ

How do I install the skill in Claude Code?

Run `/plugin marketplace add GPUtw-ai/GPUtw-Skill`, then `/plugin install gputw@gputw`. Auto-update is off by default for GitHub marketplaces, so enable it in `/plugin`, Marketplaces.

How do I connect only the MCP server?

Run `claude mcp add gputw -s user -e GPUTW_API_KEY=gputw_live_xxx -- npx -y @gputw/mcp-server@latest`. The server exposes 22 tools and takes the key from an environment variable.

How do I check the skill loaded?

Ask the agent what prefix a GPUtw API key has. The answer `gputw_live_` means the skill works.

What can I try without a key?

Open endpoints: the active GPU list `/api/gpus/active`, templates and server time. `https://upload.gputw.ai/health` also answers with no key.

Why does the API return 403 with a valid key?

There are three causes: Cloudflare blocks a request with no `User-Agent`, the key lacks the scope or the action is browser-only. The skill tells these apart first.

Discussion

No questions or reviews yet.

Skills

The official Vast.ai skill: an agent searches GPU offers, rents instances and manages them through vastai.

SkillBroad access231

Skill that powers AutoDL instances on and off from a plain-language request, keeps a budget and shuts down idle machines.

SkillBroad access95
Official

Agent skills for running, benchmarking and profiling Hugging Face models on Intel Arc GPUs, plus CUDA to XPU code migration.

PluginBroad access51

A background process manager for agents: the agent starts a dev server and reads its logs, status and ports later, plus a TUI

CLIExtended access76

A GPU profiler for AI agents: the agent runs vLLM or SGLang under graphsignal-run, reads the metrics and tunes flags for speed

SkillExtended access276
Official

The Runpod plugin with 16 skills and an MCP server: an agent starts pods, deploys endpoints and audits GPU spend

PluginBroad access49

README in short

The README is bilingual with a short English part: an example prompt, Claude Code installation, MCP-only setup and pointers for other agents. The Chinese part covers keys and scopes, the GPU catalog, instance lifecycle, live status, networking, vault, billing and teams, and debugging. The repository layout, a three-layer lookup order and safety rules close it.

README badge

Original SKILL.md

---
name: gputw
description: >
  GPUtw 台灣 GPU 雲端 REST API 整合助手（gputw, GPUtw.ai, 台灣 GPU 雲, 租 GPU, 裸機 GPU, GPU 執行個體）。
  用有權限範圍的 API 金鑰（gputw_live_）部署與管理 GPU 執行個體、輪詢狀態、讀取用量與日誌、
  在容器內執行指令、管理 HTTP 連接埠與 raw TCP/UDP、Vault 儲存空間（分段續傳上傳、Hugging Face 模型下載）、
  點數帳務與團隊。適用於撰寫自動化腳本、CI、代理程式，以及除錯 401/403/402/409/429 與 Cloudflare 403 回應。
  Use when the user mentions GPUtw, gputw.ai, renting a GPU in Taiwan, /vault, or a gputw_live_ API key.
license: MIT
metadata:
  version: "1.3.0"
  homepage: https://github.com/GPUtw-ai/GPUtw-Skill

Report a problem