Atomic Agent
A local-first AI agent running on your own machine via llama.cpp that drives a browser, edits files, runs approved commands and remembers context across session
High risk
We rate an entry high when the tool writes to external systems, handles money, production databases or secrets, or runs arbitrary commands. The CLI installs it only with your consent.
Why this level
- Runs approved shell commands, edits files and drives the browser on the user's machine
- Stores memory, sessions and provider keys locally, a risk if the machine itself is compromised
- Can optionally send data out if explicitly enabled, so egress settings should be checked
Install
Manual install
curl -fsSL https://atomicagent.io/install | shInstall on macOS and Linux.
This is third-party code. Review the repository files before installing.
What it does
Atomic Agent is a standalone terminal agent whose entire control loop and state live locally: sessions, memory, tasks, traces and the browser profile are stored as files and SQLite databases on the user's disk. On each model call it returns one grammar-checked array of tool calls and runs them without re-encoding the whole context every turn, which keeps work with small quantized models fast. It ships with a set of starter skills out of the box: Apple Calendar and Notes, Docker, ffmpeg, GitHub, Notion, Obsidian, PDF, weather and more, with external tools connecting over MCP. It includes a command to import data from Hermes, OpenClaw, Claude Code, Codex, Pi and Oh-My-Pi.
Who it is for. For developers who want a local agent without cloud token costs and full control over what leaves the machine.
Good fit when
- You need an agent that runs entirely on your own hardware on quantized models without API costs
- It matters that the agent's memory, sessions and browser profile stay only on your disk
- You have data from Claude Code, Codex, Pi or another agent that you want to migrate into a new tool
Not a fit when
- You need a stable finished tool rather than a developer preview with frequently changing APIs and commands
- There is no suitable hardware for local quantized models of a decent size
- The current Claude Code or Cursor workflow is enough and a separate local agent is not needed
Example request
Launch Atomic Agent on a local model, import my skills and sessions from Claude Code and open the browser on this pageLimitations
The README explicitly calls the project a developer preview: commands, configuration and behavior are still changing, so a stable integration should pin a release. Builds are currently available only for macOS on Apple Silicon, Linux x64/arm64 and Windows x64. Building from source requires Node.js 25.7 or higher.
How to disable. Run atomic-agent uninstall: it shows what will be removed (config, memory, sessions, models) and asks for confirmation by typing uninstall. There is a --dry-run flag to preview and --keep-data to keep your data.
Security check
- Runs approved shell commands, edits files and drives the browser on the user's machine
- Stores memory, sessions and provider keys locally, a risk if the machine itself is compromised
- Can optionally send data out if explicitly enabled, so egress settings should be checked
README in short
The README describes Atomic Agent as a local agent that drives the browser, files, commands and memory entirely on the user's machine, using its own llama.cpp build to speed up small models. It installs with a single script command, with updates and uninstall built into the binary itself. It details the agent loop: a prompt, the model deciding one grammar-checked JSON array of tool calls, execution with parallel safe actions and confirmation prompts for risky ones, and compressing results instead of accumulating the full context. It ships starter skills for Apple apps, Docker, ffmpeg, GitHub, Notion, Obsidian, PDF and more. The project is in developer preview status, MIT licensed.
FAQ
Can data be imported from Claude Code or Codex?
Yes, the first run offers an import from Hermes, OpenClaw, Claude Code, Codex, Pi or Oh-My-Pi with a dry-run preview, or the atomic-agent import <source> command later.
Is a cloud model required?
No, the agent is built around a local llama.cpp with quantized models, but cloud providers can also be connected through MCP and configuration.
Related
Open-source personal AI assistant on your own machine: answers in Telegram, Slack, Discord and WhatsApp, extended with skills and plugins
A self-improving agent from Nous Research with a TUI, messaging gateway, cron jobs and skills it writes itself
An open source coding agent for the terminal and desktop with build and plan modes
Open prompt library with a Claude Code plugin, MCP server and CLI: search, fetch and improve prompts and skills from an agent