AI Search
Ask a local OpenAI-compatible model from the GNOME Shell overview, with streaming answers and follow-up chats.
- By
- Alexander Lukens
- Publisher
- alexdlukens
- Updated
- Nov 15, 2026
- Slug
- local-ai-search-provider
Gallery
1 / 3
No releases yet
This project does not have any published versions yet.
About AI Search
AI Search puts an OpenAI-compatible model in the GNOME Shell overview. Type a question and the answer streams into the search results; open it to continue in a chat window with saved history. It is a front end only — no model is bundled, so you run a server such as Unsloth Studio, llama.cpp, Ollama, or LM Studio.
AI-generated project. The code has not been security-audited. The extension runs inside GNOME Shell with your full user privileges and can send data to the model, MCP, and web endpoints you configure.
Features
- Works with any server implementing the OpenAI chat-completions API:
Unsloth Studio, llama.cpp
llama-server, Ollama, LM Studio, or your own - Streaming Markdown answers in the overview and in a GTK chat window with saved conversations, per-question timestamps, and delete/clear controls
- Live per-turn metrics: time to first token, prompt processing speed, generation speed, and token/context usage
- Built-in read-only tools: date/time, system info, location, weather, country data, Wikipedia, currency, web search, and page fetch; optional MCP servers with a configurable tool-round cap
- Model warm-up when the overview search opens, plus loading progress when the server exposes it
- Preferences for endpoint, model, API key, thinking/reasoning, system prompt, metric or imperial units, and which tool groups are enabled
- No account, telemetry, or hosted backend. Web tools are enabled by default; turn off Enable web data tools for a fully local setup
Requirements
- GNOME Shell 50 or 51 on Wayland; older releases are not supported
- A reachable server implementing
GET <base>/modelsandPOST <base>/chat/completionswithstream: true - For source installs:
gjs,glib-compile-schemas(usuallylibglib2.0-bin), and thegnome-extensionsCLI
Install
Install AI Search from extensions.gnome.org and enable it; no build tools are required. For a source install:
git clone https://github.com/alexdlukens/local-ai-search-provider.git
cd local-ai-search-provider
make install
gnome-extensions enable ai-search@alexdlukens.github.io
make install copies the extension into
~/.local/share/gnome-shell/extensions/ and adds an AI Chats launcher
(source installs only); make uninstall removes both. Log out and back in
after installing or updating, then open the settings:
gnome-extensions prefs ai-search@alexdlukens.github.io
Set Endpoint URL to your server's base URL — a bare host:port gets
http:// and /v1 added. Leave Model on Auto, and enable tool calling
if your model supports function calling.
| Server | Typical base URL |
|---|---|
| Unsloth Studio | http://127.0.0.1:8888/v1 |
llama.cpp llama-server |
http://127.0.0.1:8080/v1 |
| Ollama | http://127.0.0.1:11434/v1 |
| LM Studio | http://127.0.0.1:1234/v1 |
Privacy
The model endpoint receives your query, system prompt, and prior turns. Web tools send their arguments to Open-Meteo, the World Bank, Wikipedia, Frankfurter, DuckDuckGo, or the page being fetched; MCP calls go to the servers you configure. API keys and MCP headers are stored as plain text in dconf and chat history is not encrypted, so keep the endpoint on loopback unless you understand the exposure.
License
GNU General Public License v3.0 or later.