Command Colab GPUs & TPUs From Your Terminal With Google’s New CLI

Google’s AI team rolled out the Colab CLI this week. This tool links your local terminal directly to remote Colab runtimes, enabling developers and AI agents to execute code on cloud GPUs and TPU hardware without ever leaving the terminal. The entire project is open source, licensed under Apache 2.0.

Introducing Google Colab CLI

Colab CLI is a command-line utility designed for Google Colab. It allows you to spin up sessions, execute code, and handle file management all from your terminal.

Any AI agent that can interact with a terminal — such as Claude Code, Codex, or Google’s own Antigravity — can leverage this tool. To make integration seamless, Google includes a pre-built instruction file called COLAB_SKILL.md, which provides agents with clear guidance on using the CLI.

Getting started is quick — just run a single uv tool install command from the GitHub repo.

uv tool install git+

Here is how a basic session looks:

colab new                              # spin up a CPU session
echo "print('hello')" | colab exec     # execute some code
colab stop                             # shut down the VM

Understanding the Commands

The CLI organizes its functionality into categories: sessions, execution, files, and automation. colab new spins up a new session, defaulting to a CPU runtime. To get a GPU, specify one with the --gpu flag — choose from T4, L4, A100, or H100. For TPUs, the available options are v5e1 and v6e1.

colab exec executes Python code fed through stdin, a .py file, or a notebook file. It reads your local files and sends their contents over automatically, so there is no need for a separate upload step every time you make a change locally. colab exec shuts down the session and frees up the VM resources when you are done.

Other useful commands handle file transfers and account access. colab upload and colab download transfer files between your local machine and the remote environment. colab drivemount connects Google Drive at the default path /content/drive. colab auth handles authentication so the VM can access various Google Cloud services.

The Core Workflow: colab exec and Recovering Artifacts

The fundamental workflow is straightforward: set up a runtime, run your script, and retrieve the results. Use colab download to pull back models, datasets, and other generated files. The colab log command saves the session history in formats like .ipynb, .md, .txt, or .jsonl.

This means every remote run can be saved as a replayable notebook right on your own machine. For hands-on debugging, colab repl and colab console offer direct interactive access to the remote VM. Need to install packages? colab install handles that with uv, falling back to pip if needed. Session data is tracked in ~/.config/colab-cli/sessions.json.

Example: Fine-Tuning Gemma 3 1B

Google’s official demo walks through an agent-driven fine-tuning process. The goal is to fine-tune the google/gemma-3-1b-it model using QLoRA on a Text-to-SQL dataset to enhance SQL generation. The Antigravity agent executes the entire pipeline with just five commands.

colab new --gpu T4
colab install transformers datasets peft trl bitsandbytes accelerate
colab exec -f finetune_run.py
colab log --output gemma_finetune_log.ipynb
colab stop

From there, the agent downloads the adapter model, adapter configuration, tokenizer configuration, and the tokenizer itself. You can load and deploy the fine-tuned model on your local machine — no manual cloud setup commands required from you.

Use Cases

• Shift training workloads from your laptop to a remote GPU or TPU, all from the terminal.
• Empower agents like Claude Code, Codex, or Antigravity to run complete ML pipelines end to end.
• Fine-tune compact models like Gemma 3 1B remotely using QLoRA.
• Automate notebook execution and export .ipynb logs for reproducible results.
• Debug interactively on the remote VM using colab repl or colab console.

Colab CLI vs. Browser-Based Colab

The CLI is not a replacement for the notebook UI — it is built for scripted, automated, and agent-driven workflows instead. Here is a side-by-side look at how the two approaches stack up.

Strengths and Considerations

Strengths:

• Terminal-native design fits naturally into scripts, CI pipelines, and agent workflows.
• T4, L4, A100, and H100 GPU provisioning with just one command.
• exec automatically sends local file contents, eliminating a separate upload step.
• Logs export to replayable notebook formats for easy reproducibility.
• Open source under Apache 2.0, with a built-in agent skill file included.
• Compatible with multiple AI agents, not locked to a single vendor.

Considerations:

• Authentication is required to access the service — the default method is oauth2.
• repl and console interactive mode requires a TTY.
• Pipe

Google Colab CLI — Terminal Guide
1 / 8

uv tool install git+

colab new                            # spin up a CPU-only session
echo "print('hello')" | colab exec   # execute inline code
colab stop                           # tear down the VM

colab new --gpu T4
colab new --gpu A100
colab new --tpu v6e1

colab exec -f train.py

colab download -s NAME checkpoints/model.bin ./model.bin
colab log -o report.ipynb

colab new --gpu T4
colab install transformers datasets peft trl bitsandbytes accelerate
colab exec -f finetune_run.py
colab log --output gemma_finetune_log.ipynb
colab stop

Marktechpost — practitioner AI & ML coverage, no hype.
Source: Marktechpost.com