What it does
Atomic Agent is an open-source, MIT-licensed agent runtime that keeps its model, agent loop and session data on your own computer. It serves a local model with llama.cpp, uses GBNF grammars to keep tool calls well formed, connects outside tools through the Model Context Protocol and stores memory in SQLite. Cloud providers are used only if you configure them. You give it a task in plain words from the terminal or Telegram, and it loops through choosing actions and running them against the browser, files and shell, pausing for approval on risky steps. The site shows example uses such as sorting and renaming files, scheduled watch-and-report checks, turning a repository issue into a pull request and researching sources into a cited note. It is available as a desktop app for Mac, Windows and Linux and as a command-line tool installed with a single command; the CLI is labelled a Developer Preview. The site also lists connections to more than 1,000 apps through Composio.
Capabilities
- Runs models locally through llama.cpp, with no accounts or token bills
- Grammar-constrained tool calls to keep actions structurally valid
- Browser, file and shell actions with approval prompts for risky steps
- SQLite-backed memory and session data stored on your own disk
- MCP support and app connections through Composio
- Task input and scheduled-check alerts through Telegram
When to use it
- You want an AI agent that keeps your data and sessions on your own machine
- You want to automate file cleanup, scheduled checks or repo issue triage with a local model
- You need an agent that can run offline or in privacy-sensitive settings
