DioxideAi

DioxideAi is a desktop LLM chat client that automatically augments every prompt with live DuckDuckGo context. Version 3 ships a revamped support dialog, per-install analytics, and a configurable Ollama or ChatGPT compatible API endpoint for remote servers.

Highlights

• Multiple chats – persistent histories with one-tap New Chat and Cmd/Ctrl + N shortcuts.
• Live streaming + cancel – models stream tokens in real time, and the send button flips to a stop button for instant cancellation.
• Deep research mode – optional multi-pass planning that gathers sources, logs progress, and inserts findings into your prompt.
• Thoughts panel – shows model load, web search, context ingestion, and timing breakdowns.
• Automatic DuckDuckGo search – fetches fresh web snippets when a prompt needs current information. Disable it in Settings for fully local replies.
• Attachment intelligence – drop .txt/.md/.json/... files with per-file and per-request limits plus truncation warnings.
• Flexible endpoints – point at local Ollama or OpenAI-compatible hosts (llama.cpp, LM Studio, etc.).
• Themes + layout – neon dark terminal look that tracks the OS or locks to light/dark, plus collapsible sidebar.
• Privacy-aware analytics – opt-in Amplitude events with detailed timing spans and hashed identifiers (see privacy.md).
• Support-friendly – in-app Buy Me a Coffee link and tutorial tips.
• Local-first – chats, attachments, and settings never leave app.getPath('userData').

Requirements

• macOS (Apple Silicon/Intel) with Node.js 18+
• Ollama running locally or remotely, or a ChatGPT-compatible server such as llama.cpp/LM Studio.
• At least one model pulled on that endpoint via ollama pull llama3:8b, ollama pull mistral, etc., or the equivalent for your ChatGPT-compatible host.

Setup

git clone https://github.com/ianrelecker/dioxideai.git
cd dioxideai
npm install
npm run dev

The dev script launches Electron with hot reload. Keep the Ollama daemon running (ollama serve) or set a remote endpoint in Settings → Connection.

Usage

1. Start the app and ensure a model is selected in the header picker (models are fetched from /api/tags on the configured host, or /v1/models when the ChatGPT-compatible toggle is enabled for llama.cpp/LM Studio endpoints).
2. Compose a prompt (drag files into the composer or use Attach file) and press Send.
3. If web search is enabled and the prompt needs current information, the app gathers DuckDuckGo snippets, streams the response, and updates the Thoughts panel as context arrives.
4. Use the Stop button to cancel long generations, open Thoughts to inspect retrieved snippets, reasoning, and timing, and toggle Hide Chats when you need a distraction-free workspace.
5. Switch endpoints or refresh the model list from Settings without restarting. Press Cmd/Ctrl + N anywhere to create a new chat instantly.

Chats auto-save and reload on launch, including canceled or failed prompts. Attachments are sent with the next prompt only (1 file, 512 KB, 1 MB total).

Settings

Open the gear icon to reveal the modal preferences panel:

• Automatic web search – opt out if you need fully offline replies.
• Max search results – number of snippets captured per prompt (1–12).
• Theme – follow macOS/Windows appearance or force Light/Dark (neon terminal mode).
• Model server endpoint – URL used for /api/tags and /api/chat (defaults to http://localhost:11434). Point this at your ChatGPT-compatible server when not using Ollama.
• llama.cpp / ChatGPT-compatible mode – makes the app talk to OpenAI-style /v1/models and /v1/chat/completions endpoints (llama.cpp, LM Studio, etc.).
• Share anonymous usage analytics – opt in/out of Amplitude tracking.
• Support DioxideAi – in-app link to Buy Me a Coffee.
• Delete all chats – wipe the history immediately.

All preferences persist at ~/Library/Application Support/DioxideAi/dioxideai-settings.json.

Packaging (macOS)

npm run dist

Electron Builder produces signed DMG/ZIP artifacts under dist/ and bundles auto-update metadata (latest-mac.yml). Configure build.publish and export GH_TOKEN (or another publisher token) before releasing. Notarise and staple the DMG if you plan to distribute outside the Mac App Store.

Tips & Troubleshooting

Issue	Fix
No models listed	Verify the configured endpoint is reachable (curl <host>/api/tags for Ollama or /v1/models when llama.cpp mode is on).
Slow first response	Large models need a warm-up load; subsequent calls are faster.
Attachments rejected	Ensure each file is ≤512 KB and the total payload ≤1 MB.
Analytics disabled	Toggle Share anonymous usage analytics inside Settings; the client honours the opt-out immediately.
Remote host change	Update the endpoint field, hit Enter (the model list refreshes automatically).

Project Structure

dioxideai/
├─ main.js         # Electron main process: Ollama + web search + persistence
├─ preload.js      # Secure IPC bridge
├─ renderer.js     # Chat UI logic & state
├─ index.html      # Renderer markup
├─ styles.css      # Renderer styles
├─ config/         # Local analytics key (gitignored) and sample template
└─ assets/         # Support graphics & app assets

License

This project is released under the PolyForm Noncommercial License 1.0.0. You may read and build upon the source for personal or internal use, but any commercial/production deployment requires a separate commercial license from Ian Relecker.