Ambient

A local-first AI desktop assistant. Run models on your machine, bring your own API keys, or use built-in cloud models — all from a lightweight app that lives in your system tray.

Built with Tauri (Rust + Next.js).

Features

Local inference

• Ships with llama.cpp and a built-in model downloader (Qwen3VL-2B)
• GPU support through the Vulkan API
• Vision support understands images and screen captures out of the box

Agentic runtime

• Web search via Tauri WebView
• Sandboxed Python code execution via RustPython
• Agentic memory across conversations
• Google Mail and Calendar integration
• Headless browser-use automation via Tauri Webview

Bring Your Own Key (BYOK)

• Connect any OpenAI, Gemini, or Anthropic-compatible API
• API keys are stored encrypted in a local SQLite database and never leave your machine

Screen context

• Snipping-tool-style region capture
• Local OCR with ocrs
• Extracted text is injected directly into the conversation

Privacy-first

• All local data (models, conversations, attachments) stays on your machine
• Encrypted databases (SQLCipher)
• Cloud features are disabled by default and opt-in

Quick Start

Prerequisites

• Rust (MSVC toolchain on Windows)
• Node.js LTS + pnpm

Development

# Install dependencies
cd app && pnpm install

# Start the Cloudflare Worker (needed for built-in cloud models)
cd cloudflare/workers/ambient-backend
pnpm install && pnpm run dev

# In a separate terminal — start the app
cd app && pnpm run tauri dev

Build

cd app && pnpm run tauri build

Architecture

app/                    Tauri application
├── src/                Next.js frontend (React 19, TypeScript, shadcn/ui)
├── src-tauri/          Rust backend
│   └── src/
│       ├── agents/     Chat + browser-use agentic runtimes
│       ├── models/     LLM providers, llama.cpp server, embeddings
│       ├── skills/     Tool registry and execution
│       ├── db/         SQLite database, migrations, CRUD
│       ├── auth/       Supabase auth, OAuth, token storage (OS keyring + AES)
│       └── settings/   User preferences
cloudflare/             Cloudflare Worker backend (cloud LLM proxy, rate limiting)

How It Works

• Local models run via llama.cpp as a background process — the app talks to it over a local HTTP server
• BYOK models connect directly to the provider's API using your key
• Built-in cloud models (Gemini) are proxied through a Cloudflare Worker
• Supabase handles authentication and user profiles
• All local artifacts (models, databases, caches, attachments) never leave your computer