AtomicBot-ai/atomic-agent

What it solves

Atomic Agent is a local-first AI agent designed to run on a user's machine to perform complex, multi-step tasks without the high costs, latency, or privacy concerns associated with cloud-hosted agents. It specifically addresses the "context bloat" problem where local models typically struggle as conversation history grows, making them slower and less effective over time.

How it works

The project implements a context-frugal agent loop that minimizes the amount of data sent to the LLM. Instead of re-sending the entire history every turn, it uses a process of prompt, decide, run, and compress, where results are summarized rather than pasted in full.

To optimize performance on consumer hardware, it utilizes a custom TurboQuant llama.cpp backend featuring KV-cache compression (up to 6.4x), Lloyd-Max weight quantization, and custom speculative decoding for models like Gemma 4 and Qwen 3.6. It also employs GBNF grammars to ensure tool calls are always valid JSON arrays.

Who it’s for

It is built for developers and power users who want a private, local AI assistant capable of driving their desktop (browser, filesystem, shell) and integrating with their own applications via HTTP or as a Tauri sidecar.

Highlights

• Desktop Control: Can navigate the web via Playwright, edit files, run approved shell commands, and manage the clipboard.
• Local-First Memory: Uses a local SQLite store for profile facts, episodic notes, and lessons, recalling them via tool calls rather than stuffing them into the prompt.
• ** uma MCP Support**: Connects to external Model Context Protocol (MCP) servers to expand its tool registry.
• Hardware Optimized: Specifically tuned for small quantized models on consumer GPUs/CPUs to maintain high throughput and accuracy.
• Flexible Providers: Supports local llama-server, OpenAI-compatible APIs, OpenRouter, and Gemini, as well as existing Claude Code and OpenAI Codex subscriptions.