nodecode-colab

Colab GPU Runner: run a Python file on a Google Colab GPU runtime and bring its outputs home. A Nodecode rewrite of PietrangeloManco/hermes-colab by PietrangeloManco.

Get it

Tell the agent "add the colab cell", or run nodecode add nc://colab.

Set it up

The cell uses your own Google account, the way the Colab web page does, and never sees your password: it needs an OAuth grant with the Colaboratory scope.

In the Google Cloud console, create or pick a project, set the OAuth consent screen to External and add your Google account under Test users.
Credentials → Create credentials → OAuth client ID, application type Desktop app. Keep its client id and secret.
Sign in once with the Google account you use on Colab and keep the grant in the colab section: either access_token (an hour), or refresh_token with client_id and client_secret, which is what lets the cell renew it by itself.

Then ask for colab_doctor — every check names the exact next step — and /colab lists the runtimes assigned to the account.

Use it

Tool	What for
`colab_run`	Run a `.py` on `T4`, `L4`, `A100`, `H100` or a CPU runtime. `accelerator`, `requirements`, `env`, `name`, `keep_runtime`, `high_mem`, `timeout_minutes`. Answers at once with a run id.
`colab_status`	One run: state, accelerator, elapsed time, outputs, events. `all` lists the recent ones.
`colab_logs`	The tail of the run's stdout and stderr.
`colab_cancel`	Interrupt the kernel, sync, release the runtime.
`colab_runtimes`	The runtimes assigned to the account.
`colab_release_runtime`	Free one by endpoint.
`colab_doctor`	The Google grant, Colab access, the active runs.

A run writes its file to /content/nodecode_colab/<run>/ on the VM, executes it there with that folder as the working directory, and brings anything the job writes under outputs/ back to ~/.nodecode/colab_runs/<run>/outputs/ while it runs; the log lands beside it in run.log. A script that reads its last checkpoint from $COLAB_OUTPUT_DIR and honours COLAB_RESUME=1 continues on a fresh runtime when Colab reclaims the VM, up to max_resumes times. Runtimes consume the account's compute units like the web page does and are released when the run ends unless keep_runtime says otherwise; check your plan's current rates in Colab.

The operator can type /colab for the connection and the runs, /colab logs and /colab cancel, and /colab <run> for one run.