Install hal0

hal0 installs from a single command. The bootstrap script is the trust boundary: it fetches the release manifest, verifies the tarball, unpacks it, and hands off to the real installer.

curl -fsSL https://hal0.dev/install.sh | bash

Note

The installer needs root to write to /etc, /usr/lib, and the systemd unit directory. When piped through bash without sudo, it re-execs itself under sudo automatically. You can also run it explicitly: curl -fsSL https://hal0.dev/install.sh | sudo bash.

Requirements

• Linux, x86_64 — hal0 ships amd64-only binaries (the FastFlowLM .deb and the toolbox container images). ARM and macOS/Windows are not supported.
• systemd — systemctl must be on PATH (outside --dev mode).
• Python ≥ 3.12 — the package requires 3.12+; the installer is tested on 3.11–3.14 and builds a dedicated virtualenv.
• A container runtime — every inference slot runs in a container. The installer auto-installs podman if no runtime is found.
• ~20 GB free disk under the data directory, plus free ports 8080 (API) and 3001 (OpenWebUI).

Supported package managers for auto-installing prerequisites (podman, the Python venv stdlib): apt (Debian/Ubuntu), dnf/yum (Fedora), pacman (Arch), zypper (openSUSE), and apk (Alpine).

NVIDIA / CUDA

hal0 is tuned for the AMD Strix Halo iGPU + XDNA NPU; ROCm is AMD-only, so GPU acceleration on other hardware runs through the generic Vulkan profile (works on AMD/NVIDIA/Intel). An experimental cuda profile also ships (the upstream llama.cpp:server-cuda image, needs nvidia-container-toolkit/CDI on the host) — it isn’t part of the one-line installer’s automatic hardware seeding, so assign it to a slot manually from the dashboard after install.

What the bootstrap does

1. Fetch the release manifest — by default the latest stable.json from GitHub Releases. Override with HAL0_RELEASES_URL or pick a channel with HAL0_CHANNEL.

2. Download and SHA-256-verify the tarball against the digest in the manifest. A mismatch aborts the install.

3. Cosign keyless verify, when available — if cosign is on PATH, the Fulcio-issued certificate and signature are checked against the workflow OIDC identity declared in the manifest; a failed check aborts the install. If cosign is not installed, the bootstrap proceeds on the SHA-256 check alone with a loud warning — cosign is no longer a hard requirement for the one-line install. Set HAL0_INSTALL_REQUIRE_COSIGN=1 to restore the old strict behavior (fail instead of warn when cosign is missing).

4. Unpack and hand off to installer/install.sh, which runs pre-flight checks, installs podman + the Python venv, builds the dashboard, writes config and systemd units, runs first-run setup (hal0 setup --auto --no-pull --no-extensions) to seed a hardware-recommended slot configuration, and starts the API.

Environment variables and flags

The installer reads several environment variables and accepts a few flags:

Variable	Default	Purpose
HAL0_PREFIX	/usr/lib/hal0	Installation root (FHS layout with a versioned dir + current symlink)
HAL0_PORT	8080	API bind port
HAL0_MODELS_DIR	/var/lib/hal0/models (or $PWD/.hal0ai/var/lib/hal0/models under --dev)	Where Hugging Face pulls land
HAL0_PYTHON	python3	Python interpreter to build the venv from
HAL0_NO_PROBE	(unset)	Set to 1 to skip first-run automatic setup (hardware-recommended slot seed)
HAL0_INSTALL_REQUIRE_COSIGN	(unset)	Set to 1 to make the bootstrap refuse the install outright when cosign isn’t present, instead of warning and proceeding on the SHA-256 check alone

Flag	Effect
--dev	Local-only editable install under $PWD/.hal0ai — no systemd, no host package changes
--no-start	Set everything up but do not enable/start the units
--models-dir=PATH	Absolute path where pulls land (same as HAL0_MODELS_DIR)

Tip

The model pull directory must be an absolute path. When not specified, the installer defaults to /var/lib/hal0/models non-interactively. You can change it later through hal0 setup.

Verify the install

Once the installer finishes, check system and slot status:

hal0 status

Re-run the full pre-flight set (architecture, systemd, Python, container runtime, disk, ports) any time:

hal0 doctor

Re-run hardware detection (CPU, RAM, GPU, NPU) on its own with:

hal0 probe

The API listens on http://localhost:8080. Service-level checks:

systemctl status hal0-api
systemctl list-units 'hal0-slot@*'
journalctl -fu hal0-api

Uninstall

The CLI wraps installer/uninstall.sh:

hal0 uninstall          # conservative: keep /etc/hal0 + models/registry
hal0 uninstall --purge  # clean slate: also wipe config, data, user, images

The conservative default stops services and removes code, units, venvs, binaries, and every hal0 podman container, but keeps /etc/hal0 (config) and the data directory (models, registry, OpenWebUI state) so a re-install reuses them. --keep-data is a legacy alias for this same default — it’s kept for back-compat, not because the default changed.

--purge (alias --clean-slate) additionally removes the config and data directories, the hal0 system user and group, and all hal0/toolbox podman images. It prompts for a typed DELETE confirmation unless you pass --force (or set HAL0_FORCE=1).

Note

--purge deliberately leaves general-purpose host packages installed — podman itself, libxrt-npu2, ffmpeg, and boost — since other software may depend on them.

Next steps

First-run setup already seeded a hardware-recommended slot and profile during install. From here:

• Run the interactive hal0 setup TUI to pick real models and extensions — see First-run setup.
• Pull a model and assign it to a slot — see Load your first model.
• Open the dashboard at http://<host>:8080/ to manage slots, profiles, and stacks.