colab-harnesslisted

Use a Google Colab GPU runtime (L4/T4, Colab Pro) as a job server from the local machine through a cloudflared tunnel, no browser automation — faster-whisper transcription, pyannote diarization, YouTube audio download, a vLLM OpenAI-compatible endpoint, LoRA training and DSPy compiles as script jobs. Use when the user asks to run something on Colab, transcribe or diarize audio on a GPU, start vLLM on Colab, train a small model on Colab, or send a batch job to Colab and fetch the results.
apresmoi/skills · ★ 2 · AI & Automation · score 73

Install: claude install-skill apresmoi/skills

# colab-harness A Colab GPU runtime becomes a job server reachable from here. Colab has no API to start a runtime, so one click stays manual (or is done through the Claude in Chrome extension); everything after that is HTTP through a tunnel. ``` you: Run all (once per session) ─→ Colab: job server + cloudflared ─→ prints the URL │ local: node colab.mjs connect <url> ←───────────────────────────────────┘ node colab.mjs pipeline / transcribe / diarize / youtube / vllm / script / run ``` All commands live in `scripts/colab.mjs`. Run `node colab.mjs --help` for flags. Each task below has one recipe; read only the one you need. | Task | Recipe | |---|---| | First-time setup: the shared token as a Colab secret | `recipes/setup-token.md` | | Gated Hugging Face models (Gemma, pyannote): HF_TOKEN and terms | `recipes/setup-hf.md` | | YouTube downloads: the cookie file | `recipes/setup-youtube-cookies.md` | | Start a session, connect, keep it alive, release it, run several | `recipes/session.md` | | Transcribe, diarize, download from YouTube, the one-command pipeline | `recipes/transcribe-diarize.md` | | Serve a model with vLLM and use it from any OpenAI client | `recipes/vllm.md` | | Run a script on the VM: LoRA training, DSPy compile, anything | `recipes/train-and-scripts.md` | ## Hard rules for an agent 1. **Never type a credential into Colab or read one off a page.** The harness token is pasted