First-run setup

First-run setup is a terminal TUI: hal0 setup. It detects your hardware, picks a coherent set of slots, models, and extensions, and provisions them — no browser required. hal0 ships with no models pre-loaded by design; setup is where you pick your first ones (or skip and configure manually later).

hal0 setup

Note

The installer runs hal0 setup --auto --no-pull --no-extensions to seed the hardware-recommended slot and write the first-run sentinel (/var/lib/hal0/.first_run_done) non-destructively — only slots whose config is absent. Re-running hal0 setup interactively afterwards lets you pull models and enable extensions. The web dashboard no longer hosts a first-run picker; it shows a banner pointing you at the terminal.

The flow

hal0 setup is a two-column terminal app: an interactive picker on the left over an always-on context pane (detected RAM, GPU, NPU) on the right. It walks:

1. Model storage — the directory every slot container reads and writes model files to. Defaults to /var/lib/hal0/models; override with --storage-dir.

2. Extensions — opt-in apps and agents, auto-wired on apply:

   Extension	Kind	Default
   Open WebUI	app	on
   ComfyUI	app (image/video, iGPU)	on
   Hermes	agent (conversational, with memory)	on
   Pi	agent (coding)	off

3. Main model — the chat model for the chat slot, picked from a short list of curated suggestions ranked to fit your detected RAM (see below).

4. Agent model — the model backing the coder slot used by agent/coding extensions (Hermes’ coding tools, Pi), from the same kind of ranked list, biased toward coder-tagged picks.

5. NPU — when an AMD XDNA NPU + FastFlowLM are present, a yes/no prompt records whether to opt in to NPU inference for this install.

Note

The Main and Agent model steps provision dedicated chat and coder slots (distinct from the agent/utility anchor roles the dispatcher and eviction logic pin by default — see Slots). hal0’s live-resolve virtuals still route correctly on a freshly-set-up box: hal0/agent falls back to a chat-named slot until you rename it or assign a model to agent directly.

Setup then applies your selection — in-process when run at install time, or via the running API (POST /api/install/apply-selections) when the daemon is up — creating the slot containers in OFFLINE and seeding background model pulls. It writes the first-run sentinel on completion.

Model suggestions

The Main and Agent model steps each show up to three curated picks — largest-that-fits-your-RAM first, with one starred as recommended — filtered to hal0’s curated model catalogue (so anything you pick already validates against the registry; nothing is invented on the fly):

Column	Shows
Model	Display name of the curated pick
Size	On-disk size in GB
Ctx	Native context length
Backend	The runtime profile it launches with (e.g. rocm-dnse, vulkan, cpu-llm)

The Agent model step ranks coder-tagged models first, falling back to general chat picks if none fit — the coder slot should never come up empty just because no dedicated coding model fits your RAM.

The NPU step (step 5) is offered independently of these suggestions, whenever AMD XDNA + FastFlowLM hardware is detected.

GPU-less hosts

On a box with no detected GPU, the cpu device now defaults to the cpu-llm profile — a real chat-capable llama-server profile — instead of one bound to the Kokoro TTS engine, so a GPU-less install still gets a working primary chat slot.

Flags

Flag	Effect
--auto	Non-interactive; apply the recommended defaults.
--storage-dir <path>	Model storage directory (default /var/lib/hal0/models).
--no-pull	Create slots + write the sentinel, but download nothing.
--no-extensions	Skip extensions (under --auto).

Configure manually instead

Prefer to wire things up yourself? Skip setup and add slots and models directly — see Load your first model.

Next

Load your first model Pull a model from Hugging Face and assign it to a slot, via setup or the CLI.