← Back to Discover

Atomic Agent

Open-source, local-first AI agent for the terminal that runs open-weight models on your machine via llama.cpp, with browser, file, shell and git tools, MCP support and memory that persists across sessions.

Autonomous agentsLocal modelsCommand-line tools
Visit website ↗

What it does

Atomic Agent is an open-source, MIT-licensed AI agent whose control loop and state run on your own computer. It uses local models through llama.cpp by default, and cloud models, or a mix of both, can be configured. It can browse the web, read and edit files, run approved shell commands, inspect documents, call external tools over MCP, schedule follow-up tasks and remember context across sessions. A command-line interface and a terminal UI cover everyday use. Approval modes range from a read-only planning mode to automatic file writes inside the workspace, and risky actions ask first. A Fusion mode lets one model plan while a pool of workers carries out the delegated parts. The README also describes remote control through Telegram and Discord, and embedding the agent in other apps over HTTP or a Tauri sidecar. Installation is one curl or PowerShell command, with builds for macOS on Apple Silicon, Linux x64 and arm64, and Windows x64. The project is marked as a developer preview, so commands and configuration may still change.

Capabilities

  • Sessions, memory and traces stored as plain files and SQLite on your own disk
  • Browser, filesystem, shell, git, document and vision tools with approval levels
  • MCP server support and Composio connections to SaaS apps
  • Local, cloud or Fusion run modes
  • Imports sessions and skills from Claude Code, Codex and other agents
  • Remote control through Telegram and Discord

When to use it

  • You want an agent that works with small quantized models on everyday hardware
  • You want a terminal agent that asks before risky actions
  • You are moving from another agent and want to bring over sessions and skills