Google Colab CLI

Use colab to inspect existing sessions or run work on an accelerator the workstation lacks. The official Colab skill documents every command; this skill owns provider selection, session hygiene, and the spend boundary.

Authentication

Pass the global --auth adc or --auth oauth2 before every subcommand: CLI 0.7.5 defaults to OAuth, so do not rely on implicit defaults or upstream main. ADC reuses credentials from gcloud; for user ADC, run dot login colab, which requests the configured ADC grant only when it lacks the Colab scopes and verifies session access (--dry-run previews the native commands; the authentication guide owns its scopes and semantics). Session state lives under ~/.config/colab-cli/. Examples assume ADC; substitute oauth2 for an existing OAuth profile.

Inspect without allocating

Check colab version and installed help, then use colab --auth adc sessions and colab --auth adc status; they synchronize session metadata without allocating or stopping a VM. Check authentication diagnostics before interpreting an empty listing as success.

Use colab --auth adc usage for compute-unit rate and balance; if it is unavailable, report a missing capability instead of allocating a VM. colab pay opens a purchase page and is not a balance query. Session inspection does not require new, run, exec, or stop.

Run accelerator work

Follow this workflow only when remote execution is in scope; establish the authorized accelerator, duration, and budget before allocation.

1. Prefer ephemeral runs for self-persisting scripts: colab run rents a VM, runs the script, and attempts to release it. Save needed artifacts to an authorized durable destination before the script exits; local VM files disappear on release. Use an explicitly managed session when artifacts must be downloaded afterwards. A shebang #!/usr/bin/env -S colab --auth adc run --gpu T4 --timeout 3600 supports single-file execution per python-script.

   colab --auth adc run --gpu T4 --timeout 3600 train.py

2. Keep a session only while iterating: colab --auth adc new -s <name> --gpu L4 (or --tpu v6e1), then colab --auth adc exec -s <name> -f snippet.py --timeout 600. Use upload, download, ls, status, and log with the same provider; verify and retrieve needed artifacts before stopping.
3. Stop what you started and verify release: colab --auth adc stop -s <name>, then refresh colab --auth adc sessions; an idle session keeps consuming compute units. CLI 0.7.5's ephemeral cleanup suppresses release errors, so a successful run exit or “Session terminated” message is not release proof. Check authentication diagnostics and the backend listing before reporting cleanup complete.

Gotchas

• Override the 30-second default: colab run and colab exec abort code execution after 30 seconds unless --timeout <seconds> covers the whole job.
• Let Colab pick the kernel client: google-colab-cli 0.7.5 pins jupyter-kernel-client==0.9.0; let Colab select its compatible client instead of adding a conflicting mise with override.
• Treat allocation side doors as spend: colab ssh creates a VM when no session exists or the -s session is missing, and colab run --keep skips the release, so both need the same authority as colab new; colab pay opens the compute-units purchase page.
• Keep the client log private: google-colab-cli 0.7.5 writes short-lived Google access tokens into ~/.config/colab-cli/colab.log (mode 0644); never read it into context, attach it, or quote it, and count matches instead of printing lines when diagnosing.
• Treat the VM as disposable: keep secrets off the session beyond what the task needs; use colab drivemount only when Drive data is required.

Official Skills

Upstream: googlecolab/google-colab-cli, the same source colab skill prints. Follow the shared vendor-skill policy and select only the Colab workflow needed by the project.

Documentation

• Colab CLI
• Releases: google-colab-cli tags
• ML workflows: python-mlops owns data validation, training, experiments, and model delivery.
• Companion skills: kaggle, hf, python-script, gcloud.