Repolode
AtomicBot-ai

AtomicBot-ai/atomic-agent

A local-first AI agent that runs its control loop and state on your machine, driving llama.cpp local models or cloud providers to browse, edit files and run approved commands.

Stars
3,415
Forks
273
Open issues
38
Last commit
2 hours ago

What is it?

Atomic Agent is a terminal and desktop-embeddable AI agent that keeps its control loop, sessions, memory, tasks and traces on the user's own machine. It works with local models served through llama.cpp, managed by the CLI or attached as an external llama-server, and can also use cloud providers such as OpenAI-compatible endpoints, OpenRouter, Anthropic and Ollama.

The agent can drive a Playwright-based browser, read, write and patch files, run shell commands behind an approval system, extract text from PDF and Office documents, and connect MCP servers. Memory is stored in SQLite with full-text and optional embedding-based recall, and scheduled tasks, webhooks and Telegram or Discord bots let it run unattended.

It exposes an OpenAI-compatible HTTP endpoint, a newline-delimited JSON stdio sidecar for Tauri apps, and a TUI with approval modes, a plan mode and multiple color themes. A Fusion mode lets a cloud orchestrator delegate work to local worker models.

Key features

  • 01Drives a browser, edits files and runs approved shell commands.
  • 02Runs local GGUF models through llama.cpp, or cloud providers when configured.
  • 03Stores sessions, memory and config as local files and SQLite databases.
  • 04Provides an OpenAI-compatible HTTP API and a Tauri sidecar protocol.
  • 05Connects to Telegram and Discord bots and to external MCP servers.
  • 06Schedules recurring tasks and reminders with cron or intervals.

Use cases

  • Running a coding or browsing agent on a laptop with a small local model and no API costs.
  • Embedding an agent in a desktop app through the stdio sidecar protocol.
  • Controlling a home or work agent from Telegram or Discord while it runs on your own machine.
  • Automating recurring file or web tasks with scheduled jobs and approval gates.

Quick start

  1. 1

    Run the official install script from the README (macOS/Linux or Windows PowerShell)

  2. 2

    Launch the agent

    atomic-agent

Things to know

  • Developer preview with APIs, commands and config still changing; pin a release for stable integration.
  • Intel Macs are not supported yet.
  • Installs via a curl-to-shell style script rather than a package manager.

Topics

FAQ

What is atomic-agent?
A local-first AI agent that runs its control loop and state on your machine, driving llama.cpp local models or cloud providers to browse, edit files and run approved commands. Atomic Agent is a terminal and desktop-embeddable AI agent that keeps its control loop, sessions, memory, tasks and traces on the user's own machine. It works with local models served through llama.cpp, managed by the CLI or attached as an external llama-server, and can also use cloud providers such as OpenAI-compatible endpoints, OpenRouter, Anthropic and Ollama.
Is atomic-agent free to use? What license does it have?
atomic-agent is open source under the MIT license. Check the license terms before using it in commercial products.
Is atomic-agent still maintained?
Actively maintained. The last commit to atomic-agent was on Oct 22, 2026.
What can I use atomic-agent for?
Running a coding or browsing agent on a laptop with a small local model and no API costs. Embedding an agent in a desktop app through the stdio sidecar protocol. Controlling a home or work agent from Telegram or Discord while it runs on your own machine. Automating recurring file or web tasks with scheduled jobs and approval gates.
How do I install atomic-agent?
Run the official install script from the README (macOS/Linux or Windows PowerShell) Launch the agent: atomic-agent
Where can I get atomic-agent?
The source code and installation instructions are on GitHub: https://github.com/AtomicBot-ai/atomic-agent

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects