What it does

MiniSearch is a self-hosted metasearch engine that queries the web through SearXNG, then uses an AI assistant to synthesize answers. The twist: the LLM can run entirely in your browser via WebGPU, so your queries never hit a remote model server unless you want them to. You can also plug in your own OpenAI-compatible API on the backend and let others use it without exposing your key.

The interesting bit

The project treats the browser as the compute layer. Models load and cache on demand via WebGPU, which means desktop and mobile users alike get local inference without installing anything. It’s a neat inversion of the usual “send everything to the cloud” pattern.

Key highlights

• Runs in a single Docker container; docker run and you’re live
• Browser-based LLM inference with GPU acceleration (WebGPU)
• Privacy defaults inherited from SearXNG: no tracking, no ads, no data collection
• Address-bar integration: set http://localhost:7860/?q=%s as your default search
• Optional “internal API” mode lets you share your own API key without revealing it
• Customizable search and generation settings through the UI

Caveats

• Requires Docker; no bare-metal install path is documented
• WebGPU support varies by browser and platform, so local inference may not work everywhere
• The README doesn’t specify which models are bundled or their size, so bandwidth and memory requirements are unclear

Verdict

Good fit for privacy-conscious developers who want a drop-in replacement for their default search engine and don’t mind self-hosting. Skip it if you need enterprise audit logs or heavy customization beyond the exposed settings.