All commands

Every hi command and option in one place.

Overview

hi compute

Options for up, run, and serve

Defaults: --gpu is the provider’s cheapest CPU; --max is 4h for up and serve and 1h for run, at most 24h on Colab. Flags may come before or after names for logs, serve, stop, and billing.

hi connect and hi server

See Managed compute for your team.

hi data

Needs a connected server with organizations added, and hf (pip install -U huggingface_hub). The Hugging Face token stays on the server; hf runs with HF_ENDPOINT pointing to it.

Put dataset:, model:, or bucket: in front when several share a name. Buckets have no revisions. A group’s data patterns in policy (["hifinab/*"]) limit what it sees; without them it sees everything.

hi login

hi login hf       # hf auth login, then hf auth whoami
hi login colab    # the Colab CLI's browser sign-in, then the balance
hi login runpod   # asks for the API key, checks it, saves it in ~/.runpod/config.toml
hi login shadeform # asks for the API key, checks it, saves it in ~/.config/hi/shadeform_key

hi init

hi init                          # choose a template, name, and directory
hi init python pricing-tools     # create ./pricing-tools from the python template
hi init web . --name dashboard   # apply the web template to this directory
hi init --list                   # the templates
hi init --update                 # bring this repository up to the current templates
hi init --update --check         # change nothing; fail when it is behind (for CI)
hi init --adopt python           # bring this existing repository under a template

Templates: python, web, service, pipeline, ml, and on a connected device your team’s private ones.

hi q

Everything from the first word that doesn’t start with - is the question, so hi q status of the log file asks the model. hi’s own options go before it; put -- before a question that starts with a dash.

Options: --provider server|openai|openrouter|anthropic|claude and --model <name> for one question or chat, --no-context to send only the question, and --yes to run read-only commands without asking. Dangerous commands always need yes typed. See Ask for a command.

hi agent

Options: --detach to start and return, --json for the report as JSON on stdout, --task-file <path> for a task file whose name doesn’t end in .md, --model <model> for the agent’s model (without it, the one in your own Claude Code, Codex, or Hermes settings, else opus for Claude Code), --bundle a,b for bundles of skills and the tools they need, and the box options below. A brief’s front matter can set agent, model, bundles, network, allow, data, and gpu. Exits with 0 when the agent is done and 1 when it failed. See Hand tasks to agents.

best-of also takes --agents claude,codex (taken in turn), --check "<command>" (default: make check), --max <duration> per agent, and --yes to start without asking; it exits with 1 when no box passed. With rounds: --for <duration>, --budget <dollars>, --patience <n>, --min-gain <number>, --edit a,b for the files an agent may change, and --then "<command>", run on this machine when the rounds end. To keep a gain only if it also holds on data agents never see: --validate "<command>" (run by hi after each round in a container without network or .git, with only the --edit files from the box; needs --edit), --validate-data <dir> (a folder outside the project, read-only at /box/data there), and --validate-slack <number> (how much worse than the highest validation so far the number may be; default 0); a winner that fails is rejected. --folds <n> --fold "<command>" walks the validation window forward through n folds, after --fold-every <k> rounds or when --patience runs out; each fold starts from a new baseline. See Keep only gains that hold. See Autoresearch.

hi team

See Run a team in Slack.

In its box, the lead runs hi team task <role> <brief.md> [--approval <message id> | --small | --on <task>] [--attach <file>]..., review <task>, wait <task>, show <task>, tasks, merge <task>, stop <task>, env (the names of the app’s settings), channel [--since 2h | --new], web (the chats and feedback in the app), web show <chat|feedback>, web post <chat> <text|-> [--attach <file>]..., web questions (what people asked the app’s help), web link --user <member ID> (DM someone their private links, as /hi does) (a message in a chat after its answer), and preview [<task> | production] [--port <n>] [--max <duration>] or preview stop, which runs the project in an app box and puts it on a temporary address with hi net expose. versions shows production’s version and what dev has on top of it; release --approval <id> --notes <file> puts dev into production as the next version once an owner says yes, and release v2 --approval <id> puts production back on v2; discard <task> drops an idea that isn’t merged, and discard dev --approval <id> sets dev’s code back to production’s.

hi box

Options when starting: --name, --network locked|dev|open, --allow <domain>, --gpu, --data (your team’s data through the hi server), --worktree, --here, --image, --memory, and --bundle a,b. A project’s devcontainer.json can set network, domains, gpu, data, and bundles under customizations.hi. See Run code in a box.

hi bundle

Bundles come from hi itself, your team’s template sources on a hi server, and ~/.local/share/hi/bundles (or HI_BUNDLES_DIR); a later one wins on a name. See Bundles.

hi skill

Options: --global (your home folder), --skill a,b, --yes, --accept-risk, --ref, --force, and --json. hi skills is the same command. See Skills.

hi install and hi verify

hi install                     # menu: check to install, uncheck to remove
hi install --list              # tool names, and which are installed
hi install claude gh           # install or reinstall the named tools
hi install --all [strix]       # every tool; Strix Halo support only when named
hi install strix               # the menu with Strix Halo support checked
hi uninstall docker            # remove the named tools
hi verify strix                # the Strix report again, after a reboot

hi net

hi net                                   # enroll; prompts for the key
hi net --setup-key-file <file>           # enroll from a file only you can read
HI_NETBIRD_SETUP_KEY=… hi net            # enroll from the environment
hi net status | down | reconnect

See Share a local service.

hi adduser

hi adduser <name>    # Ubuntu adduser, plus the render and video groups