Quickstart

Sign in with GitHub

Open the dashboard and Sign in with GitHub. We request the repo scope so the fixing pipeline can read your training repo and open pull requests on your behalf.

Create an API key

Go to API Keys → New key. Copy the ka_live_… value — it is shown once. This key authenticates the library when it reports metrics and events.

Create a Project

Go to Projects → New Project and pick the GitHub repo that holds your training code. We store the repo and install a push webhook on it. Merging to its default branch is what launches training.

Install and log in

On your machine:

pip install keepalive-club
keepalive login   # paste your ka_live_ key; stored in ~/.config/keepalive/config.json

Set WANDB_API_KEY in your environment too — training reports to Weights & Biases.

Wrap your training loop

Add a watchdog context manager around your loop. The prompt is plain English: describe what healthy training looks like for this model.

import keepalive, wandb

run = wandb.init(project="my-model")

with keepalive.watchdog(
    run,
    prompt="Flag if val/loss diverges from train/loss or grad norms spike",
    threshold=0.6,
    max_agents=3,
):
    train()

The watchdog hooks run.log(), batches metrics and events to the dashboard, and runs hard-failure detectors (NaN loss, OOM, stalls, exceptions) in-process with zero lag.

Merge to main

Push your wired-up training script and merge to your default branch. The push webhook fires the training launch workflow: we spin up a W&B Sandbox, clone your repo at the merged commit, pip install -r requirements.txt, and run your train command. Branch pushes never trigger training — only the default branch does.

Watch the dashboard

Open the project's Overview. You'll see the live loss curve, the event feed, and the run status. When the monitoring agent flags something — or a hard failure trips — an Incident appears, the fixing pipeline spawns coding agents, and PRs land on your repo, each with a written report.