Seedbox
Privacy-first local LLMs, with a polished UI
Consumer AI chat apps send every message to someone else’s cloud. Users who care about privacy need a local-first alternative — but wiring Ollama, a vector DB, and a polished desktop UI together is not a weekend project.
Seedbox packages the whole stack into a single installable app: pick a local model, talk to it, keep everything on your machine.
Chat as a graph, not a transcript
Built the conversation surface on React Flow (Xyflow). Instead of a linear scroll of messages, chats live as nodes on a canvas — you can branch, compare responses side-by-side, and navigate threads spatially. Long exploratory sessions feel structural, not endless.

Local-first data, end to end
SQLite for chat history, LMDB for vector embeddings, optional Redis for caching. Nothing leaves the device.
A Fastify server runs inside the Electron main process, exposed to the renderer through a custom `app://` protocol — the UI gets the ergonomics of a web app while every byte stays local.

Ollama integration with a settings surface
Model selection, system-prompt config, and sampling controls surfaced through a settings panel. Llama 3.2 as the default; any Ollama-compatible model plugs in from the UI.

Details that add up
A Three.js animated background so the app doesn’t feel like a stock Electron shell. Light/dark theme with persistent preference. Markdown with math (KaTeX), GFM tables, and remark directives — so technical answers render properly. None of this is the feature, but all of it is what makes the tool feel considered.

Cross-platform builds
electron-builder configs for Windows (NSIS installer), macOS (DMG — code-signing pipeline ready), and Linux (AppImage / Snap / Debian). One codebase, three distributions. Seedbox has shipped alpha → beta → alpha-release; public on GitHub as source-available.