Atomic Agent is a local-first AI agent that runs open-weight models on your own machine via llama.cpp.
Atomic Agent addresses the need for autonomous AI that operates entirely under your control, without sending data to external services. It runs a complete agentic loop locally: the agent can browse the web using browser automation, read and edit files, execute approved shell commands, inspect documents, and maintain context across sessions. The tool supports both local models and cloud models, with a Fusion mode that lets one model handle planning while a pool of workers executes tasks. It can be embedded in your own applications over HTTP or as a Tauri sidecar, and integrates external tools through the Model Context Protocol. The project uses llama.cpp with TurboQuant quantization to keep small models performant for long, multi-step workflows on consumer hardware.
Atomic Agent suits developers who want an AI agent without cloud dependencies or data transmission concerns, particularly those working on local machines with modest resources. It works well for automation tasks spanning file editing, web interaction, and command execution where you need to approve actions before they run. The tool is in developer preview, so APIs and behavior remain in flux; adopting it requires pinning a specific release if you need stability. The project offers data import from several other agent platforms, letting you migrate skills, memory, and sessions without starting fresh.
The project maintains active development with regular self-update capability built into the CLI. The installer verifies checksums and manages native prebuilds alongside grammar files and bundled tools. The team supports multiple platforms including macOS with Apple Silicon, Linux on x64 and ARM64, and Windows x64, though Intel Macs are not yet supported. Documentation covers development setup, and the tool includes a terminal UI with commands for onboarding, window management, and data import workflows.
