SSH-ing in to tail the log
From your phone, between classes. The card already shows the epoch and the metric.
For researchers, grad students & labs
Step away from
your experiment.
Training runs, simulations, sweeps and long data pipelines take hours. Have your agent watch the run and keep a live card on your iPhone with the epoch, the metric you care about, and when it should finish. Go to class, go to sleep, go get lunch.
Free for one agent and one live card · Works with Claude Code, Codex and Grok
Why
The run is long. Checking it shouldn't be.
You don't need a dashboard. You need to know, at a glance, whether to walk back to the lab.
A dashboard tab
Great for analysis, overkill for “is it still going?” The Lock Screen answers that.
Waking up to a crash
Hours lost. The agent can mark the card alert the moment the loss goes NaN or the job dies.
How it works
Your agent watches. You glance.
1
Connect your agent
Add Startr to Claude Code, Codex or Grok once, and approve it on your iPhone.
2
Start the run through the agent
Ask the agent to launch the job (or attach to one) and report from the log: epoch, metric, ETA.
3
Glance, don't check
The card updates in place on your Lock Screen and Apple Watch, and ends with the final result.
One night, four glances
From launch to result without opening a laptop.
Illustrative cards.
◆ lab-agentSTARTED
Fine-tune run · sweep 3Loading data · 2.1M examples
ETA 6h 20m0 / 40
◆ lab-agentRUNNING
Fine-tune run · sweep 3Val loss 0.214 · best so far
ETA 2h 10m14 / 40
◆ lab-agentWARN
Fine-tune run · sweep 3Val loss rose 3 epochs in a row
ETA 1h 30m22 / 40
◆ lab-agentDONE
Fine-tune run · sweep 3Best val loss 0.198 at epoch 31
Total 6h 51m40 / 40
Copy these to your agent
Prompts for long jobs.
Training run
Run train.py and watch its log. Keep a Startr card updated each epoch with the epoch count, validation loss and an ETA. Mark it warn if validation loss rises three epochs in a row, alert if the job crashes, and end it with the best result.
Simulation or sweep
Launch the parameter sweep in sweep.sh. Keep a Startr card with runs completed out of total and the best score so far, and end it with a one-line summary.
Data pipeline
Run the pipeline in pipeline/. Use a Startr journey card with the waypoints Download and Clean and the destination Analysis, and update it at each stage.
Good to know
What it does, and what it doesn't.
Show what matters to you.
Epoch, loss, accuracy, runs done, rows processed. The agent writes the card from whatever the log says.
Lock Screen, Dynamic Island, Watch.
The card sits where you already look. No app to open, no tab to refresh.
Your data stays yours.
Startr only receives the short card text. It never sees your code, data, or results files.
Quiet means quiet.
If the agent stops updating, the card goes stale after two hours instead of pretending the job is still running.
The agent has to stay running to keep the card current.
Startr doesn't watch your job by itself: the agent session sends each update, so keep that session alive (for example in tmux on the machine running the job). Each card lasts up to eight hours; for longer runs, ask the agent to start a fresh card before the first one ends.
Questions
Before your next run
Does my cluster need anything installed?
No. The agent reads your job's output and talks to Startr. Nothing runs inside your job.
Can it watch a job that's already running?
Yes, if the agent can read the job's log or status command. Point it at the log file.
Does it work with Slurm or remote machines?
If your agent can run commands where the job lives (for example over SSH), it can report from there.
Is it free for students?
The free tier covers one agent and one live card, which is enough for one run at a time.
Android?
Not today. Startr is an iPhone app.
Agent Cards for research
Launch it.
Leave it.
Set it up once, then every long run reports to your Lock Screen.
Read the research guideRunning something unusual? Tell us, we'll help you set it up.