Google Colab CLI
A tiernew this weekRun Python on remote Colab GPUs and TPUs straight from your terminal — no browser, no notebook babysitting required.
Kai's verdict
Genuinely useful unlock for anyone already in the Colab ecosystem — turns a browser-bound notebook service into a proper programmable compute backend that agents can actually drive. Hardware availability limits and Windows exclusion keep it from being an instant essential, but the agent-native design is a smart differentiator. (Verdict pending Phi's full review.)
Strengths
- Instant provisioning of T4, L4, A100, H100, and TPU v5e/v6e runtimes with a single command
- Agent-native design — ships a COLAB_SKILL.md that lets Claude Code, Codex, or any terminal agent drive full ML workflows autonomously
- Ephemeral job runner (colab run) handles VM spin-up, script execution, artifact retrieval, and teardown in one shot
- Works with existing Colab quota and credentials — no new cloud account, no manual infrastructure setup
- Supports local Python scripts, .ipynb notebooks, stdin pipes, interactive REPLs, and Google Drive mounting
Weaknesses
- Linux and macOS only — Windows users are completely locked out at launch
- GPU/TPU availability still subject to Colab's dynamic quota system, so you can't guarantee hardware on the free tier
- Very new and openly still in development — rough edges and reliability issues should be expected
Best for
ML engineers and AI agent builders who want cheap, scriptable access to Colab GPUs without ever opening a browser or managing cloud infrastructure.
Pricing
Free (CLI is open source) + Colab Pro $11.99/mo or Pro+ $49.99/mo for premium GPUs
The CLI itself is free and open source (Apache-2.0). Accelerator access is gated by your Colab subscription: free tier gets limited T4 access; A100, H100, and TPUs require Pro or Pro+.