Atomic-agent is a local-first AI agent designed for robust desktop work, prioritizing privacy and user control by running entirely on the user's device, even offline. It optimizes local AI models, supports long context windows via Turboquant, and features proper tool calling, making it suitable for users who want an AI agent they can fully own and inspect, moving beyond simple chat interfaces to real-world task automation.

How It Works

The agent employs a purpose-built backend, atomic-llama-cpp-turboquant, tuned for throughput on consumer hardware, and utilizes curated quantized models. Its core innovation lies in efficient context management: a stable prefix for persona and tools, combined with aggressive KV-cache reuse and bounded prompt tails, minimizes re-encoding. Tool calls are constrained via GBNF grammar, and inferences produce JSON arrays of tool calls that are executed, results compressed, state updated, and the loop repeats, ensuring cheap and valid inferences for multi-step tasks.

Quick Start & Requirements

• Install: curl -fsSL https://api.atomicbot.ai/agent-install | sh
• Run: atomic-agent
• Prerequisites: Node.js (for development), a reachable llama-server (managed or external), Chromium-family browser (Chrome, Edge, etc.), git. Linux requires desktop tools like ripgrep, clipboard utilities (xclip/wl-clipboard), libnotify-bin, wmctrl (X11/XWayland), and gio/trash-cli. GPU acceleration requires Vulkan drivers.
• Platform: macOS and Linux x64. Windows builds are forthcoming.
• Docs: Core documentation includes PROMPT.md, MEMORY.md, SKILLS.md, among others.

Highlighted Details

• Benchmarks on the GAIA L1 split show atomic-agent achieving 69.8% accuracy and ~217s per task, outperforming Hermes (58.5% accuracy, ~351s/task) under identical conditions.
• Designed for local models, it features a stable prefix for KV-cache reuse and bounded prompt tails, keeping inference costs low.
• Offers extensive desktop capabilities including browser automation, filesystem/shell access, document parsing (PDF, DOCX, XLSX, etc.), Git operations, and a flexible skill system.
• Features an externalized, inspectable memory system using SQLite + FTS5, embeddings, and recall mechanisms, avoiding large prompt dumps.

Licensing & Compatibility

The project is licensed under the MIT license, permitting broad compatibility for commercial use and integration into closed-source applications.

Limitations & Caveats

This project is in Developer Preview, with APIs, commands, and behavior subject to change. Current releases are limited to macOS and Linux x64; Windows support is pending. Browser sandboxing on some Linux setups may require the --no-sandbox flag. Secret redaction and isolation layers are not yet complete.