Claw Code Agent

Overview

Claw Code Agent is a Python reimplementation of Claude Code’s agent architecture¹ designed to run entirely with local open-source models via OpenAI-compatible APIs. The project achieves full agentic coding workflows — prompt assembly, context building, tool execution, session persistence, and nested delegation — in pure Python with zero external dependencies.

Where Claude Code requires Anthropic’s proprietary API, Claw Code Agent targets local inference backends like vLLM, Ollama, LiteLLM Proxy, and OpenRouter, with first-class support for Qwen3-Coder-30B-A3B-Instruct as its recommended model.

Key Findings

• Implements a complete agentic loop mirroring Claude Code: prompt assembly, iterative tool calling, streaming output, context compaction, and budget enforcement — all in ~21,000 lines of pure Python across 51 source files.
• Ships 60+ built-in tools covering file operations, search, shell execution, MCP integration, task management, team collaboration, git worktree management, and nested agent delegation.
• Achieves strong parity with Claude Code’s core execution layer (agent loop, tool calling, streaming, session persistence, cost tracking) while lacking ~70% of the npm implementation’s interactive UI, remote infrastructure, and service ecosystem.
• A sophisticated bash security system with 18 validators and 163 tests detects injection attacks, obfuscation, destructive commands, and parser-differential exploits before shell execution.
• Plugin system supports manifest-based extensions with lifecycle hooks, tool aliases, virtual tools, and tool-level interception — enabling customization without modifying core code.
• MCP (Model Context Protocol) runtime implements JSON-RPC 2.0 over stdio transport for resource discovery and tool invocation through external servers.
• Zero external dependencies — the entire codebase uses only the Python standard library (Python 3.10+).

Motivation

Claude Code is a powerful agentic coding tool, but it is locked to Anthropic’s API and closed-source infrastructure. Claw Code Agent asks: what if you could run the same architecture locally, with any model, and full visibility into the internals? The project is aimed at developers who want:

• Full control over model selection, inference parameters, and cost
• Privacy — code never leaves the local machine
• Extensibility — modify any part of the agent pipeline
• Education — a readable reference implementation of agentic coding patterns

Footnotes

References

• Claw Code Agent GitHub Repository
• Claude Code Documentation
• Qwen3-Coder Model
• vLLM Project
• Model Context Protocol Specification

Contents