Connect Aether to Ollama

60-second setup

Already ran ollama pull … once? Then you are one step away. This is the exact path for a Windows machine with Ollama installed and models already on disk:

Start the Ollama app Open it from the Start menu. It lives in the system tray (you may need to click the ^ chevron). Ollama runs as a background server — closing the window keeps it running, quitting the tray icon stops it.
Confirm it answers In any terminal: ollama list should print your models. If the list works, the server is up — the CLI talks to the same server Aether uses.
Open Aether's AI sidebar Click the ✨ sidebar icon in the toolbar (or press Ctrl+Shift+A). Within a moment the status dot turns green and your models appear in the dropdown next to it — picked automatically, no configuration screen.
Ask something Chat, AI search answers and the agent all use this model now. Done.

Saw the red dot first? Since v1.4.1 Aether auto-reconnects: while Ollama is offline it re-checks every 10 seconds and whenever the window gets focus. Start Ollama at any time — Aether connects by itself within seconds. No restart needed. On older versions, click the ⟳ button or Retry now.

How the connection works

Aether never phones home — it talks straight to the Ollama HTTP server on your machine:

Piece	Detail
Endpoint	`http://127.0.0.1:11434` — the Ollama default. Override with the `AETHER_OLLAMA_URL` environment variable (see remote hosts).
Health check	`GET /api/tags` with a short timeout. This is what the status dot shows — same call the `ollama list` CLI makes.
Auto-detection	On connect, Aether reads your installed models and picks a default: it prefers Qwen-family models with tool calling, falling back to the first model on disk. You can override the choice in the dropdown at any time — the pick is saved.
Auto-reconnect	While the dot is red, Aether re-checks every 10 s and on window focus. Ollama coming online is detected without any click.
Where it runs	In the main process (Node), not the web view — so there are no CORS or proxy issues, and requests never leave your machine.

One server, everything shares it. The AI sidebar (Chat / Agent / Models), AI search answers and the omnibox all resolve through the same local brain. Select a model once and it applies everywhere.

Use your own model (fine-tunes, GGUF, anything)

Aether works with any model Ollama serves — including your own fine-tuned GGUF from Hugging Face. Say you have this library:

C:\Users\you> ollama list
NAME                                              ID            SIZE      MODIFIED
coder:latest                                      100e5c7ae56f  4.3 GB    3 days ago
hf.co/mayankbhaskarr/qwen3-4b-coder-ft-gguf:latest  557cfa5d33cd  4.3 GB  3 days ago

The moment Aether connects, both appear in the model dropdown in the sidebar header. Pick one and it is stored as your brain:

Switch models — dropdown in the sidebar header. Applies to chat, AI search and the agent immediately.
Models tab — Ctrl+Shift+M opens the full model manager: details (parameters, quantization, disk size), delete, and the in-app library where you can pull any tag from Ollama or Hugging Face without a terminal.
Agent quality tip — the agent depends on tool calling. Qwen-family models handle it well; a 4B fine-tune works, but for heavy multi-step automation a 7–9B model gives noticeably steadier step planning.
First answer is slow? That's Ollama loading the model into RAM/VRAM (a 4.3 GB model takes a few seconds). Follow-up answers stream instantly while the model stays loaded.

Pulling a new model in-app: Models tab → type the exact tag (hf.co/user/model-GGUF, qwen3.5:4b, …) or pick a catalog card → Pull. Progress streams live; the model appears in your dropdown the moment it finishes.

Verify it works

Three checks, ten seconds:

The dot is green Green = Ollama reachable and a local model selected. Yellow = Ollama offline but an optional cloud fallback is answering chat. Red = neither (see troubleshooting). Hover the dot for the exact status.
Your model is in the dropdown Open the header dropdown — every entry is a real model on your disk, straight from ollama list.
Send a message In Chat, ask anything — the reply streams token by token from your machine. Or open aether://search, search something and watch the cited AI answer build on the right.

Extra confirmation: run ollama ps in a terminal right after an Aether reply — you'll see your model listed as loaded, with Aether as the recent caller.

Remote or custom Ollama host

Want Aether to use Ollama running on another machine (a GPU box, a home server, WSL2)? Point the AETHER_OLLAMA_URL environment variable at it before starting Aether:

# Windows (PowerShell)
$env:AETHER_OLLAMA_URL = "http://192.168.1.50:11434"; npm run dev

# Windows (CMD)
set AETHER_OLLAMA_URL=http://192.168.1.50:11434 && npm run dev

# Linux / macOS
AETHER_OLLAMA_URL=http://192.168.1.50:11434 npm run dev

On the machine running Ollama, expose it beyond localhost first:

# Server side — allow LAN connections
OLLAMA_HOST=0.0.0.0 ollama serve     # or set it in the service config

WSL2 note — Ollama inside WSL is reachable from Windows Aether at http://127.0.0.1:11434 only with WSL's mirrored networking; otherwise bind it and use the WSL IP.
No Ollama anywhere? Aether also accepts any OpenAI-compatible endpoint (LM Studio, Groq, OpenRouter) as an opt-in cloud fallback for chat and AI search — Settings → AI → cloud fallback. The agent always stays local.
Security — only expose Ollama to networks you trust; it has no authentication.

Troubleshooting

Red dot, but `ollama list` works in the terminal

This is the most confusing state — and the answer is usually timing. The CLI succeeding proves the server is up right now, so Aether's last check simply ran before that. The check happens at sidebar open; Ollama may have still been starting after boot (its first-load can lag a few seconds on Windows). Fixes, in order of convenience:

Wait ~10 s with the Aether window focused — auto-reconnect (v1.4.1+) will flip the dot green by itself.
Click the ⟳ button in the sidebar header, or Retry now in the red hint box.
Make sure the Ollama tray app is running, not just a terminal — ollama list in a terminal starts a temporary CLI that talks to the server; it does not start the server. Start-menu → Ollama starts the background server.

“Ollama is not reachable” persists

Open http://127.0.0.1:11434 in any browser — it should say “Ollama is running”. If not, the server truly isn't up: start it from the Start menu, or run ollama serve.
Different port? Check echo %OLLAMA_HOST% (CMD) / $env:OLLAMA_HOST (PowerShell). If you moved it, set AETHER_OLLAMA_URL to match (e.g. http://127.0.0.1:11500).
Proxy software (Clash, corporate proxies) can intercept localhost calls for Electron apps. Add an exclusion for 127.0.0.1 / enable "bypass local addresses".

Model missing from the dropdown

Press ⟳ to refresh the list; the Models tab always shows the live inventory too.
Model pulled while Aether was already running? The Models tab and the dropdown refresh automatically the moment a pull finishes.
Typo'd tag in a manual pull? The tag must match exactly what ollama list shows, including the :latest suffix.

Answers are very slow

First call after boot always includes model load time — subsequent replies are fast while it stays warm.
A 9B model on a machine without a GPU can take a while per token; a 4B model (qwen3.5:4b, your 4B fine-tune) is dramatically quicker on CPU.
Check what else holds VRAM — another loaded model can force a reload every time. ollama ps shows what's resident.

The agent errors but chat works

The agent needs tool calling, not just chat. Some small models ignore the tool format. Pick a Qwen-family model for agent runs — the chat models will still handle chat, search and write helpers fine.