Summary

Claw Code Agent is a pure Python reimplementation of the Claude Code agent architecture, designed for local execution with open-source LLMs. It offers engineers and power users a zero-dependency, fast-iteration environment for complex coding and agentic workflows, providing full control over models and data.

How It Works

This project leverages Python's standard library exclusively, avoiding external package dependencies. It emulates the Claude Code agent's core loop, including prompt assembly, context management, tool calling, and session persistence. The agent interacts with local LLM inference servers (like vLLM, Ollama, LiteLLM Proxy, or OpenRouter) via an OpenAI-compatible API, enabling local model execution and full data control.

Quick Start & Requirements

• Primary Install/Run: Execute via python3 -m src.main agent <prompt> or agent-chat.
• Prerequisites: Python 3.10+ is required. A local LLM server supporting tool calling is essential, such as vLLM (configured with --enable-auto-tool-choice --tool-call-parser qwen3_xml for Qwen3-Coder), Ollama, LiteLLM Proxy, or OpenRouter. The Qwen/Qwen3-Coder-30B-A3B-Instruct model is recommended.
• Links: vLLM Tool Calling Docs, Ollama OpenAI Compatibility, LiteLLM Docs, LiteLLM Proxy Quick Start, OpenRouter Docs, OpenRouter Models, OpenRouter API Keys.

Highlighted Details

• Features an extensive set of agentic capabilities, including interactive chat, token-by-token streaming, a manifest-based plugin runtime, nested agent delegation with topological batching, and robust cost/budget tracking.
• Supports structured output via JSON schema, file history replay, and context compaction techniques for efficient prompt management.
• Offers broad LLM backend integration, working seamlessly with Ollama, LiteLLM Proxy, OpenRouter, and vLLM, providing flexibility in model selection.
• Includes advanced runtimes for local search, MCP transport, task/plan management, remote execution, hook/policy enforcement, and git worktree operations.

Maintenance & Community

The project is developed by the HarnessLab Team. No specific community channels (e.g., Discord, Slack) or detailed contributor information are provided in the README.

Licensing & Compatibility

The repository explicitly states "Zero external dependencies — just Python's standard library." However, no specific open-source license (e.g., MIT, Apache) is mentioned. Users should verify licensing terms for commercial use or integration into closed-source projects.

Limitations & Caveats

Several features are marked as "In Progress," including full MCP parity, slash-command parity, REPL/TUI behavior, and real remote transport. The project is a reimplementation and not affiliated with Anthropic. Using cloud-based gateways like OpenRouter requires caution due to data transmission.