guide

Let your coding agent use a GPU

Give Claude Code, Codex, or omp the hi skill so they can run your code on rented GPUs - asking before they spend, and cleaning up after.

Coding agents can run hi like any other command. The hi agent skill teaches them how, and the rules to follow: ask you before spending, preview with --dry-run, always set --max, and stop what they start.

1. Add the skill to your project

cd my-project
hi skill add hi

Wrote /home/you/my-project/.agents/skills/hi/SKILL.md
Wrote /home/you/my-project/.claude/skills/hi -> ../../.agents/skills/hi
Agents started in this folder now know how to use hi. Commit the files to share them.

Commit both paths so everyone’s agents get the skill. For every project on your machine, use hi skill add hi --global instead. Skills has the details, and how to add other skills from skills.sh.

2. Ask for what you want

Start your agent in the project and ask in plain words:

Run train.py on a GPU with 24 GB of memory for 200 epochs and tell me the final loss.

Serve Qwen3-8B on a cheap GPU for half an hour so I can test my app against it.

Is anything still running on Colab or Hugging Face? Stop what you started.

3. What the agent does

Following the skill, the agent:

  1. Chooses hardware from hi compute hardware and runs the command with --dry-run.
  2. Tells you the provider, hardware, price per hour, maximum lifetime, and on Hugging Face who pays, and waits for your yes.
  3. Runs the command with --yes. Agents have no terminal for hi’s own confirmation, so --yes is how your approval reaches hi.
  4. Uses --secret for credentials and never prints tokens.
  5. Stops what it started and checks hi compute ls when done.

The skill tells agents never to add --yes without your approval. If your agent is allowed to run commands without asking, you are still protected by --max, but review its plans for paid hardware.

Keep the skill up to date

After updating hi, run hi skill again in each project, or hi skill --global, to refresh it with the new commands.