The AI assistant

W ships an AI assistant that runs in a terminal and can drive the machine through W's own tools — read the display layout, switch a power profile, look at logs. It is optional in every sense: nothing depends on it, and it does nothing until you configure a model for it.

Super+W opens the ask palette; typing a question there hands off to a full session. w-ai in a terminal is the same thing, and w-ai status prints what is configured right now.

Profiles #

Hub → AI is a list of profiles. A profile is a saved combination of which program runs the assistant, which model it talks to, and how much it may do without asking. Tapping an inactive profile makes it active immediately; the active one carries a checkmark.

The point of profiles is switching setups without editing anything: keep one for a cloud model and one that runs entirely on your own machine, and flip between them in a tap. New profile… creates one, optionally copying an existing one as a starting point.

The chevron on a row expands the editor. Every change applies as you pick it — there is no save button — and none of it asks for your password: this is a personal setting, like a keyboard layout.

Host, provider and model #

• Host is the program that actually runs the session. W ships with one included; the others are third-party CLIs with their own login and their own updater, so they are installed per user, on request.
• Provider is the service behind the model. What the dropdown offers adapts to the host you picked — for a CLI that has its own subscription login it offers subscription, which means W passes no API key at all and the vendor bills your plan rather than per token.
• Model is the model id. It is required for the included host — leave it empty and the session will not start. For a provider CLI it is marked optional, because that CLI picks its own.

When the chosen host needs something, a row appears right underneath it saying so — not installed with an Install button, or not signed in with Sign in. Both open a terminal and run the step for you. Installing a host does not need your password: it lands in your own account.

If a session refuses to start, w-ai ready in a terminal answers why in one word.

The third-party CLIs W can drive are Claude Code, Codex and OpenCode. W hands them its tools and its identity only when it launches them itself — start claude, codex or opencode by hand in a project of your own and you get the plain tool, with nothing of W attached. OpenCode is the one with its own subscription (Zen) and a direct line to dozens of other providers at once — pick subscription for Zen, one of the providers W already manages a key for, or set up anything else straight in OpenCode itself (opencode auth login).

Some packs bring tools of their own for the assistant — the graphics pack, for instance, lets it draw in Inkscape and edit inside GIMP, and the 3d pack lets it build scenes in Blender. These arrive the same way, at launch, for every host — with Goose they stay switched off until a request needs that app, so everyday questions don't carry their weight; w-ai status lists them under MCP extra. One shown as missing is installed on the machine but not set up for your account yet: w-pack setup <pack> (no password) and a fresh session.

API keys #

Where the provider needs a key, the expanded profile shows an API key row: whether one is stored, a masked field to paste one into, and a button to remove it. The key goes into the system keyring, never into a file on disk, and it belongs to the provider rather than to the profile — every profile using that provider shares it. From a terminal the same thing is w-ai key set.

One trap worth naming: if you are using a CLI on its subscription login, do not also store an API key for it. With a key present the CLI switches to per-token API billing.

Mode: how much it does on its own #

Mode is chat, approve, smart approve or auto, in increasing order of independence: from answering only, through asking before each action, to acting on its own.

This is a convenience setting, not the security boundary. Anything that touches the system as administrator stops at W's own password or fingerprint prompt no matter which mode you picked — auto cannot authorise a privileged change on your behalf. For the same reason W's own tools (including the ones an installed pack adds, like the Inkscape or GIMP helpers) never wait on the host's per-tool confirmation: the assistant calls them directly, and anything privileged still stops at that same prompt. MCP servers you register yourself keep the host's normal confirmation flow.

Tool surface next to it controls how much of W's own tooling the assistant is offered. Leave it on auto.

Advanced #

The Advanced section adds capabilities that need extra software: Web search, Reader (fetching and reading a page), Semantic memory (searching its own memory by meaning rather than by keyword) and Local models.

These come from an optional bundle. Until it is installed the four rows are greyed out with Requires the ai-extra pack and a single Install button that does it for you — see the packs guide. Afterwards each row switches on independently, and the ones W cannot enable for you report why: that the local model runner is not running, or that a model still has to be downloaded. Choosing brave for web search reveals its own API-key field, which works exactly like the one above.

Nothing here is required for an ordinary chat session; it is the difference between an assistant that can look things up and one that can only answer.

ComfyUI #

Generation of images and video is a separate optional bundle — see the packs guide. It runs ComfyUI as a system service: a node-based editor and backend combined, on http://127.0.0.1:8188/. Open it from the app menu (which starts the service, asks for the admin password the first time and opens the browser) or start it yourself with systemctl start comfyui.

• GPU matching: the engine is installed with the driver your hardware needs (nvidia / amd / intel Arc / cpu), detected automatically.
• Adding nodes: the node manager is built in and already switched on, so a node pack you are missing is installed from the interface itself — nothing to clone, nothing to install by hand. It only works while ComfyUI answers on 127.0.0.1: expose the engine to the network and the manager stops installing anything, which is its own safety rule rather than W's.
• Shared model store: models are downloaded into /var/lib/w/ai-models, a snapshot-excluded store shared by the whole machine; every account set up for the pack writes there. Ollama from the ai-extra pack keeps its models there too, in ollama/.
• Where your results land: /var/lib/w/comfyui/data/output (under the engine directory, wherever it was installed). You can open that folder in the file manager, collect renders and drop source images into input/ next to it — access comes with the w-ai group the bundle adds you to, which takes effect at your next login.
• Updating: a system update leaves the engine alone; w-pack outdated shows a newer release and sudo w-pack upgrade comfyui installs it, putting the old one back if the new one does not start — see keeping packs current.
• The assistant can help with ComfyUI workflows, model placement and the service — load the comfyui knowledge pack when asked about local generation.

InvokeAI #

A second, independent generation bundle with a different interface: InvokeAI works on a layered canvas (inpainting, outpainting, regional prompts) with a gallery and boards, where ComfyUI builds a node graph. It runs as a system service on http://127.0.0.1:9090/; open InvokeAI from the app menu or run systemctl start invokeai.

• GPU matching: the PyTorch build is chosen for your hardware at install (nvidia / amd / intel Arc / cpu), the same way as for ComfyUI.
• Same models as ComfyUI: every model file in the shared store (/var/lib/w/ai-models/checkpoints, loras, vae, …) is registered in InvokeAI where it lies — nothing is copied. Added a file? Run sudo systemctl reload invokeai and it appears after a short check. Files InvokeAI cannot use are skipped quietly.
• Its own downloads from the Model Manager tab go to /var/lib/w/ai-models/invokeai. Removing a model inside InvokeAI deletes a file only from there; a shared-store model is just hidden until next start.
• Where your results land: /var/lib/w/invokeai/data/outputs, reachable from the file manager with the same w-ai group (next login).
• Tuning (memory limits, attention, precision) lives in /var/lib/w/invokeai/data/invokeai.yaml; the example file next to it lists every setting.
• Updating: the same as ComfyUI — sudo w-pack upgrade invokeai, which also says up front whether the new release brings a new PyTorch (a large download) and restores the previous version and its database if the new one fails.

Edit this page on GitHub