AI queries that never leave your machine.

Pick a provider

Open the gear icon in the viewer, then the AI query tab. Enable the providers you want and drag them into order — a query uses the first enabled one that's available on your machine. All three local options run with nothing sent anywhere: your file, your question, and the model all stay on the device.

• Ollama — points QDev at a local Ollama server on localhost:11434. Strongest local models, but you install and run it yourself.
• Local (WebLLM) — runs a model inside the browser tab over WebGPU. Downloads it on first use (roughly 2.5–5 GB depending on the preset) and caches it for next time. Needs a Chromium-based browser.
• Chrome built-in— uses the small model Chrome ships when the browser provides it. Nothing to configure or download, but it's the weakest of the three.

Set up Ollama

Install Ollama from ollama.com/download, pull a model, and start the server so a browser page is allowed to reach it:

ollama pull qwen2.5-coder  # or any model you prefer

OLLAMA_ORIGINS="*" ollama serve  # start it with browser access allowed

That's it. While the Ollama section of the settings panel is open, QDev re-checks the server every couple of seconds, so the status flips to Detected and reachable on its own once the server is up. The endpoint and model fields are there if you run Ollama somewhere other than localhost:11434 or want a different model than qwen2.5-coder.

“Running but blocking browser extensions”

Ollama only answers browser requests from origins it trusts, and extension pages aren't trusted by default — so a plain ollama serve rejects QDev with a 403 even though the server is running fine. When that happens the settings panel says so explicitly instead of showing the provider as unavailable. The fix is one environment variable on the server side:

OLLAMA_ORIGINS="*" ollama serve  # quickest — allow any origin

OLLAMA_ORIGINS="chrome-extension://*,moz-extension://*" ollama serve  # scoped — extensions only

On macOS, if you use the menu-bar Ollama app rather than running ollama serve in a terminal, set the variable for the app and then quit and reopen it:

launchctl setenv OLLAMA_ORIGINS "*"  # then quit and reopen Ollama

Allowing an origin only controls who may talk to the server on your machine — your files and questions still never leave it.

Which one fits

Local models trade convenience and accuracy for privacy and price — WebLLM needs a real download and WebGPU, the built-in model is small, and Ollama wants capable hardware. See how the four AI modes compare and where each local option hits a wall for the honest breakdown.