Humans and machines, creating together.

Install the workstation software. Offers a hostname change, then asks for sudo once.

Everything above plus ROCm and GPU groups, finishing with a local hardware report.

Re-run the hardware report after the reboot. Nothing is uploaded.

Enroll with NetBird. The setup key is read without echo and never lands in shell history.

Check, disconnect or reconnect an enrolled peer without a new key.

Put a web app or webhook receiver on your machine, even on localhost, on a temporary public HTTPS address. Password-protected by default, ends on its own, logs every request.

Start a project agents work well in: make check, short agent rules, safe settings and CI. Python, web, service, pipeline, ML and autoresearch (model training) built in, plus your team's own, and --update keeps projects current.

Ask for a shell command in plain words, or chat. The model looks at your files first; you see the command, and what its globs match, before Enter runs it. Uses Claude Code, OpenRouter or any OpenAI-compatible endpoint.

Hand a task to Claude Code, Codex, or Hermes. It works without permission prompts in a box, on its own branch, and you get one report back, the same for every agent: what it said, whether it worked, and what changed. --bundle web,office,data adds skills and tools for browsing, documents and data.

Run one task in n boxes at once, mixing agents. hi runs your check in each box itself and ranks the results, and you keep one. With --score it runs rounds, autoresearch-style: each starts from the best so far, and hi commits a result only when its score beats it. --then runs a command of yours, such as a report, when the rounds end.

A virtual team of agents in one Slack channel: a Hermes lead talks with your people and keeps the specs, a coder and a reviewer work in boxes, and the app runs in previews and production. hi team key gives it an OpenRouter key of its own and hi team env the app's settings, so nobody pastes a key into Slack.

Run a shell or a command in a rootless container: the project and nothing else, no credentials, a network that reaches only what you allow, and the Strix Halo GPU if you want it. devcontainer.json sets it up for the whole project.

Guided menu to start, reach and stop remote GPU machines on Colab, Hugging Face, RunPod or Shadeform, with lifetime limits.

Pick a running machine from a list, or all of them, and stop it after confirming.

Run a script or container to completion on rented hardware and stream its logs.

Serve a GGUF model with llama.cpp on a rented GPU and tunnel its OpenAI-compatible API to localhost.

Join your team's hi server, which approves and pays for GPUs and serves your team's project templates.

Search your team's private Hugging Face datasets, models and buckets and download one with hf, or read them straight from code with hi data run. The server keeps the token; your machine never holds one.

Give a cloud job your team's data: the server opens a narrow public address with NetBird for the run, so the job downloads it and can call load_dataset itself, with no Hugging Face token on the machine.

Find, install, and update agent skills from skills.sh, pinned and audited, and teach agents to use hi safely.

Create a user who already has render and video access.

Update hi itself to the latest release, checksum-verified. --check only looks.

Print the installed version.