Anthropic's terminal agent, reassembled in Python

What it does Clawd Code is a Python reimplementation of Anthropic’s Claude Code terminal agent. It reconstructs the TypeScript tool-calling loop, streaming REPL, and skill system as a working CLI that supports multiple LLM providers. The project ships with 30+ built-in tools, session persistence, and a tab-completing interactive shell.

The interesting bit Instead of treating Claude Code as a black box, the authors ported the real TypeScript architecture into idiomatic Python, then wrapped it in a markdown-driven skill system where slash commands are just SKILL.md files with tool sandboxes. It is explicitly an independent educational project, not an official Anthropic product.

Key highlights

• Supports Anthropic, OpenAI, and Zhipu GLM out of the box, with a provider-agnostic backend.
• 30+ tools covering file operations, Bash execution, web fetch/search, task management, and MCP integration.
• Skills are programmable via markdown frontmatter: define arguments, descriptions, and allowed tool lists in a SKILL.md.
• Core test suites exist for skills, providers, REPL, tools, and context.
• Claims startup under one second and memory under 50 MB.

Caveats

• Context building and the permission system are marked as partially implemented; deeper project understanding is still listed as needed.
• Roadmap phases 4 and 5 (MCP plugins, Python-native differentiators) are pending, so extensibility beyond the current skill system is not yet realized.

Verdict Worth a look if you want a readable, hackable Python reference for how a modern coding agent loop works. Skip it if you need a battle-tested, fully permissioned IDE integration today.