Claude Code, but make it 233,000 lines of Python

What it does ClawCodex is a from-scratch Python port of Anthropic’s Claude Code, translated from the TypeScript reference implementation into roughly 233,000 lines of idiomatic Python. It preserves the full terminal agent experience—streaming REPL, multi-turn tool loops, session persistence, and slash-command skills—while letting you swap LLM providers at will.

The interesting bit This is not a thin API wrapper. The codebase includes a remote-control bridge, a dynamic workflow engine with agent()/parallel()/pipeline() primitives, a /deep-research harness, and even a /buddy virtual-pet command. The scale implies they actually ported the architecture, not just the network calls.

Key highlights

• Multi-provider runtime: Wires Anthropic, OpenAI, DeepSeek, GLM, MiniMax, and OpenRouter into the same tool-calling loop, so you can route to cheaper or stronger models without forking the agent.
• /advisor cost mode: Pairs a cheap worker model with an expensive reviewer consulted only at decision points; the README cites roughly 6× cost savings versus running the expensive model solo.
• Benchmark edge: On the full SWE-bench Verified split using Gemini 2.5 Pro, it resolved 58.2% of issues versus openclaude’s 53.0%.
• Markdown skills: Custom slash commands defined via SKILL.md files with typed arguments and tool allow-lists.
• Workflow engine: Supports parallel agents, worktree isolation, and live TUI monitoring; the /deep-research tool even forbids tool access during its final synthesis step to prevent runaway loops.

Verdict A solid candidate if you want a hackable, Python-native coding agent that refuses to be locked to one model family. Avoid if you need a minimal, drop-in script—this is a full-stack application with the gravity to match.