guide

The hi guide

Step-by-step guides for hi, the hi5 command-line tool. Set up a workstation, then rent GPUs on Colab or Hugging Face to run jobs, open shells, and serve models.

hi does two jobs. It sets up hi5 workstations, and its hi compute commands rent remote machines so you can use a GPU you do not have: run a training script, open a shell, forward a port to your laptop, or serve a model with an OpenAI-compatible API. The same commands work on Google Colab and Hugging Face.

Where to start

What you can do

I want to… Read
Run a Python script or container on a GPU Run a job to completion
Get a shell on a GPU machine and forward a port Interactive machines
Try an LLM on hardware I do not have Serve a model
Show a web app or receive a webhook on my machine Share a local service
Download your team’s private datasets and models Download your team’s data
Give a cloud job your team’s data Run a job: your team’s data
Know what I am paying and who pays Costs, limits, and billing
Set up Colab or Hugging Face for the first time Google Colab, Hugging Face Jobs
Hand a task to Claude Code, Codex, or Hermes in a sandbox Hand tasks to agents
Try a task several times and keep the best result Several attempts at once
Let agents improve a number overnight, keeping only gains Rounds that keep only gains
Start an autoresearch project for a model Start from a template
Let Claude Code or Codex use a GPU safely Let your coding agent use a GPU
Run a team of agents in a Slack channel Run a team in Slack
Turn “move the md files to notes” into a command Ask for a command
Start a new repository that agents work well in Start a project
Set up a new hi5 machine Set up a workstation
Fix an error message Troubleshooting

Conventions

Commands are shown as you type them. Replace names in angle brackets, such as <name>, with your own. hi help lists every command, and hi compute help lists the compute commands and their options.

Commands that rent hardware cost money. Each one prints the price before it starts (and on Hugging Face, who pays), and asks you to confirm paid hardware. Add --dry-run to see what would happen without starting anything.

Press / anywhere to search the guide.