SafeRL-Lab/cheetahclaws

What it solves

CheetahClaws is a Python-native agent harness designed to provide a flexible, open-source alternative to proprietary coding assistants like Claude Code. It solves the problem of being locked into a single model provider or a compiled, hard-to-modify codebase, allowing developers to run AI agents with any LLM (including local models via Ollama) and customize the agent's tools and behavior at runtime.

How it works

The project implements a streaming API and an automatic tool-use loop (contained primarily in agent.py) that enables an LLM to interact with a system. It uses a multi-layered context compression system to manage token limits and a dual-scope persistent memory system to track user and project information. The harness supports a wide array of built-in tools for file manipulation, shell execution, and web searching, and can be extended via a plugin system or the Model Context Protocol (MCP).

Who it’s for

• Developers who want a local or non-Anthropic AI coding assistant in their terminal.
• Researchers studying the inner workings of agentic AI systems.
• Teams needing a hackable, readable baseline for building custom AI agents without a Node.js build chain.

Highlights

• Multi-Model Support: Compatible with Anthropic, OpenAI, Gemini, DeepSeek, and local models via Ollama, LM Studio, or vLLM.
• Extensible Tooling: Includes 28 built-in tools and supports runtime tool registration, MCP integration, and git plugins.
• Advanced Context Management: Features four layers of context compression and a task dependency graph for long-horizon tasks.
• Agentic Capabilities: Supports spawning typed sub-agents (e.g., coder, reviewer), autonomous background loops, and specialized modes like /research and /trading.
• Developer Experience: Provides a terminal REPL with next-prompt ghost text, a Web UI, and bridges to messaging platforms like Telegram and Slack.