{"id":95771,"date":"2026-06-07T17:39:02","date_gmt":"2026-06-07T17:39:02","guid":{"rendered":"https:\/\/youzum.net\/googles-new-colab-cli-lets-developers-and-ai-agents-run-python-on-remote-colab-gpus-and-tpus-from-the-terminal\/"},"modified":"2026-06-07T17:39:02","modified_gmt":"2026-06-07T17:39:02","slug":"googles-new-colab-cli-lets-developers-and-ai-agents-run-python-on-remote-colab-gpus-and-tpus-from-the-terminal","status":"publish","type":"post","link":"https:\/\/youzum.net\/th\/googles-new-colab-cli-lets-developers-and-ai-agents-run-python-on-remote-colab-gpus-and-tpus-from-the-terminal\/","title":{"rendered":"Google\u2019s New Colab CLI Lets Developers and AI Agents Run Python on Remote Colab GPUs and TPUs From the Terminal"},"content":{"rendered":"

This week, Google AI team released the Colab CLI<\/a><\/strong>. The tool connects your local terminal to remote Colab runtimes. It lets developers and AI agents run code on cloud GPUs and TPUs. You stay in your terminal the entire time. The CLI is open source under the Apache 2.0 license.<\/p>\n

What is Google Colab CLI<\/strong><\/h2>\n

The Colab CLI is a command-line interface for Google Colab. You can create sessions, run code, and manage files from the terminal.<\/p>\n

Any agent with terminal access can call the tool. That includes Claude Code, Codex, and Google\u2019s Antigravity. Google ships a prepackaged skill file named COLAB_SKILL.md<\/code>. It gives agents built-in context on how to use the CLI.<\/p>\n

Installation uses a single uv tool install<\/code> command from the GitHub repository.<\/p>\n

\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
uv tool install git+https:\/\/github.com\/googlecolab\/google-colab-cli<\/code><\/pre>\n<\/div>\n<\/div>\n
A minimal session looks like this:<\/strong><\/p>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
colab new                              # provision a CPU session\necho \"print('hello')\" | colab exec     # run code\ncolab stop                             # release the VM\n<\/code><\/pre>\n<\/div>\n<\/div>\n
How the Commands Work<\/strong><\/h2>\n
The CLI groups commands into sessions, execution, files, and automation. colab new<\/code> provisions a session, with CPU as the default. Add --gpu T4<\/code>, --gpu L4<\/code>, --gpu A100<\/code>, or --gpu H100<\/code> for a GPU. TPU options are v5e1<\/code> and v6e1<\/code>.<\/p>\n
colab exec<\/code> runs Python from stdin, a .py<\/code> file, or a notebook. The exec<\/code> reads files locally and ships their contents. Local edits therefore need no separate upload step. colab stop<\/code> terminates the session and releases the VM.<\/p>\n
Other commands cover files and authentication. colab upload<\/code> and colab download<\/code> move files between local and remote. colab drivemount<\/code> mounts Google Drive, defaulting to \/content\/drive<\/code>. colab auth<\/code> authenticates the VM for Google Cloud services.<\/p>\n
colab exec<\/code> and Artifact Recovery: The Core Loop<\/strong><\/h2>\n
The core loop is short. You provision a runtime, run a script, then pull results back. colab download<\/code> retrieves models, datasets, and other files. colab log<\/code> exports session history as .ipynb<\/code>, .md<\/code>, .txt<\/code>, or .jsonl<\/code>.<\/p>\n
So a remote run becomes a replayable notebook on your disk. colab repl<\/code> and colab console<\/code> give interactive access to the VM. colab install<\/code> adds packages with uv<\/code>, falling back to pip<\/code>. Session metadata is stored at ~\/.config\/colab-cli\/sessions.json<\/code>.<\/p>\n
Example: Fine-Tuning Gemma 3 1B<\/strong><\/h2>\n
Google\u2019s official release demonstrates an agent-driven fine-tuning job. The task fine-tunes google\/gemma-3-1b-it<\/code> using QLoRA. It trains on a Text-to-SQL dataset to improve SQL generation. The Antigravity agent runs the full pipeline with five commands.<\/p>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
colab new --gpu T4\ncolab install transformers datasets peft trl bitsandbytes accelerate\ncolab exec -f finetune_run.py\ncolab log --output gemma_finetune_log.ipynb\ncolab stop<\/code><\/pre>\n<\/div>\n<\/div>\n
The agent then downloads the adapter model, adapter config, tokenizer config, and tokenizer. You can load and serve the fine-tuned model locally. No manual cloud provisioning command was typed by the user.<\/p>\n
Use Cases<\/strong><\/h2>\n