The AI Hub
The AI Hub is the desktop side of everything AI. It has three tabs — AI Agents, Knowledge Base and Settings — and it is where the work that is awkward in a browser gets done: setting up the environment, downloading multi-gigabyte models, and pointing the index at folders.
First run: the three-step wizard
A machine that has never run AI gets a wizard rather than a wall of controls:
- Environment — checks for Python 3.10 or newer and installs the dependencies. This is the step that takes minutes and downloads the most.
- Model — the list is matched to this computer: its processor, memory, graphics card and free disk space are compared with an up-to-date list of models, and the best fit is preselected. Each row says where the model would run (graphics card, graphics card + RAM, or processor) and how fast.
- Launch — start the service and load the model.
If the environment check reports Python is missing, install it and run the check again; on Windows the installer normally provides a private copy, so this usually only appears with the portable build. See running Filemote without installing it.
AI Agents
Main Orchestrator is the model that answers by default. Its card shows whether the service is online or stopped, whether weights are still loading, and a request count.
- Start AI / Stop AI control the service.
- Change Model or Replace swap the loaded model.
- Unload frees the memory a model is holding without stopping the service.
- Model Catalog expands to every model that fits this computer, for the main model or for a specialist slot. Each download is checked against a fingerprint before it is used. Typing in the search box also searches HuggingFace, for picking a main model by hand.
Before starting, the configurator can validate a model: it checks the model exists, that it is a chat-capable type, and that the machine has enough memory — with a plain warning like "model needs ~7 GB, only 4 GB available" rather than a failure ten minutes later.
Specialist agents are the document, vision, code and summary slots. Loading one routes that kind of question to it; leaving it empty falls back to the orchestrator. Each loaded model holds memory for as long as it is loaded, which is the whole trade-off.
About the Stop button
Stop AI stops the AI service, and it is worth knowing that this is now true in a way it was not always.
The configurator and the agent can both own the same AI service on port 8765. There was a version where Stop only ended the handle the window happened to hold, so a service started by the other side kept running — and a status poller would notice it alive and effectively bring it back. If you have ever seen the AI "start itself", that was it, and it is fixed: Stop ends the process on the port and records that you asked for it, so nothing revives it behind your back.
AiAutoStart is a separate setting: it starts the AI service with the agent, and it is not
affected by pressing Stop.
Knowledge Base
The counters at the top show chunks, documents and the size on disk.
Indexed Folders is the list that gets read; + Add Folder and Remove manage it, and indexing shows progress with a Stop of its own. Two extra sources sit behind consent checkboxes: Browser history and Email (.eml/.msg) — how far each one travels is set out in AI models and the knowledge base.
A useful detail: knowledge-base work does not need the full AI service. When the agent is not running, the configurator starts a small knowledge-base-only helper, so indexing a folder does not mean loading a language model first.
Settings
Model defaults, the compute device, reply length, the system prompt, and the indexing
schedule. Every field here maps to one in app_settings.json, listed in
configure the agent by hand.
When it will not start
The Hub reports what happened rather than failing silently — a missing model, an unavailable server, a failed load. Each message names the step. The list of causes, in the order worth checking, is in the AI does not load.
Updated 2026-10-02