Launching Claude Code without an Anthropic key 🏴‍☠️

A very interesting project called free-claude-code has recently appeared. It's essentially a proxy server on FastAPI that replaces Anthropic endpoints. You feed it requests from the original Claude Code, and it routes them wherever you want: to the free NVIDIA NIM tier, to OpenRouter, directly to DeepSeek, or even to a local LM Studio / llama.cpp.

What's interesting here (besides the obvious freebie):

1️⃣ Heuristic Tool Parser
Cheap or locally run quantized models often struggle with calling tools in a strict format. The author wrote their own state machine (providers/common/heuristic_tool_parser.py) that parses the raw text output of the model and wraps it into a legitimate Anthropic tool_use block on the fly. An ideal crutch for dumb models to make them work with Claude's tools.

2️⃣ Interception of service traffic
In api/optimization_handlers.py, local interception of "junk" requests is implemented. Under the hood, Claude Code constantly sends requests for quota checks, title generation for history, autocomplete of file paths, and suggestions. The proxy handles these itself with hardcoded responses, without ever calling the external LLM. Saves time, money, and limits.

3️⃣ Real-time reasoning translation
ThinkTagParser extracts <think> tags (hello, DeepSeek-R1) or reasoning_content from the provider's response in real time, mapping them to official Anthropic thinking blocks. To the client, it looks like a native Claude is responding.

As a wild bonus: the proxy has handlers for Telegram and Discord. You can point a bot at a local project folder, send it voice messages (transcribed via local Whisper), and the agent will write code on your computer. Perfect if you want to fix production while riding in a taxi (please don't do that).

Try it now! Let us know if another saving grace works ☕️