Free Open-Source Proxy Unlocks Unlimited Claude Code Usage

Background

Claude Code is Anthropic’s terminal‑level AI programming assistant that can understand whole codebases, perform Git operations and multi‑file refactoring. Its official API is expensive, difficult to access from China, and locked to Anthropic‑provided models.

free‑claude‑code

The open‑source project free‑claude‑code (GitHub https://github.com/Alishahryar1/free-claude-code) gained >20 000 stars within a week and provides a lightweight proxy that forwards Claude Code requests to free or low‑cost LLM providers while preserving the Anthropic API contract.

Core principle

Implemented in Python/FastAPI, the proxy intercepts Claude Code’s HTTP calls, routes them to a selected back‑end (NVIDIA NIM, OpenRouter, DeepSeek, LM Studio, llama.cpp, Ollama), converts the provider’s response to the Anthropic format and returns it unchanged to Claude Code. No modification of Claude Code’s CLI, VS Code or JetBrains extensions is required.

┌─────────────────┐   ┌──────────────────────┐   ┌──────────────────┐
│  Claude Code    │──►│  Free Claude Code    │──►│  Real LLM vendor │
│  CLI / VSCode   │◄──│  Local proxy (:8082)│◄──│  NIM / OR / local│
└─────────────────┘   └──────────────────────┘   └──────────────────┘
   Anthropic API format          OpenAI‑compatible format (SSE)

Key features

Support for six back‑ends covering free, paid and self‑hosted models.

Intelligent model routing: Opus, Sonnet and Haiku requests can be directed to different providers.

Native /model endpoint compatibility – Claude Code can switch models without extra configuration.

Full feature compatibility: streaming, tool calls, reasoning blocks, with latency optimisations.

Optional Discord or Telegram bot, voice‑note transcription via Whisper or NVIDIA NIM.

Provider overview

NVIDIA NIM – free, 40 req/min, suitable for daily use; requires an nvapi‑… key.

OpenRouter – free or paid, rate limits vary by model.

DeepSeek – pay‑as‑you‑go, Anthropic‑compatible API.

LM Studio – free local, unlimited rate, privacy‑focused.

llama.cpp – free local, unlimited rate, lightweight inference.

Ollama – free local, unlimited rate, easiest to start.

Technical highlights

Thought‑chain preservation – parses reasoning_content tags so models such as DeepSeek R1 or GLM display their reasoning in Claude Code.

Automatic tool‑call repair – heuristics fix non‑standard tool‑call outputs from open models.

Smart rate‑limit and anti‑ban – concurrency control and 429 back‑off reduce latency and prevent bans.

Advanced usage

Discord or Telegram bots can turn Claude Code into a remote‑controllable agent with threaded conversations, persistent sessions and voice transcription. The project also ships a claude‑pick script that uses fzf for interactive model selection, and a voice‑note feature configurable via VOICE_NOTE_ENABLED and Whisper or NVIDIA NIM models.

Quick start

Install the base dependencies:

# macOS/Linux (bash)
curl -LsSf https://astral.sh/uv/install.sh | sh
uv self update
uv python install 3.14

or

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
uv self update
uv python install 3.14

Clone the repository and copy the example env file:

git clone https://github.com/Alishahryar1/free-claude-code.git
cd free-claude-code
cp .env.example .env   # or Copy-Item on Windows

Edit .env to add the API key of the chosen provider (example for NVIDIA NIM) and model mappings:

# NVIDIA NIM key (required)
NVIDIA_NIM_API_KEY="nvapi-YOUR_KEY"
# Model routing
MODEL_OPUS="nvidia_nim/moonshotai/kimi-k2.5"
MODEL_SONNET="nvidia_nim/z-ai/glm-5.1"
MODEL_HAIKU="nvidia_nim/z-ai/glm4.7"
MODEL="nvidia_nim/z-ai/glm4.7"
# Optional features
ENABLE_MODEL_THINKING=true
ANTHROPIC_AUTH_TOKEN="freecc"
PROVIDER_RATE_LIMIT=1
PROVIDER_RATE_WINDOW=3
PROVIDER_MAX_CONCURRENCY=5

Start the local proxy:

uv run uvicorn server:app --host 0.0.0.0 --port 8082

Configure Claude Code to point to the proxy (CLI, VS Code or JetBrains): set ANTHROPIC_BASE_URL=http://localhost:8082 and ANTHROPIC_AUTH_TOKEN=freecc. Example for the CLI:

ANTHROPIC_AUTH_TOKEN="freecc" ANTHROPIC_BASE_URL="http://localhost:8082" claude

and for VS Code add the same variables to settings.json.

{
  "claudeCode.environmentVariables": [
    {"name": "ANTHROPIC_BASE_URL", "value": "http://localhost:8082"},
    {"name": "ANTHROPIC_AUTH_TOKEN", "value": "freecc"}
  ]
}

Troubleshooting

Incorrect response format – ensure ANTHROPIC_BASE_URL has no /v1 suffix and the proxy returns streaming data.

Local model returns 400 – verify the local service is running, supports /v1/messages, and increase context size if needed.

Streaming interruptions – lower concurrency, increase timeout parameters, or switch to a more stable provider.

VS Code shows login screen – check settings.json environment variables and that the proxy is running.

Limitations

Because the back‑end model is swapped, tasks that require very long context, multi‑step tool calls or deep code refactoring may perform worse than the original Anthropic models. For everyday coding, single‑file edits and code explanations the free models are sufficient and cost‑effective.

Project structure

free-claude-code/
├── server.py          # ASGI entry point
├── api/               # FastAPI routes and logic
├── core/              # Protocol utilities and streaming
├── providers/         # Provider adapters and rate limiting
├── messaging/         # Chat platform adapters and voice
├── cli/               # Command‑line entry
├── config/            # Configuration and logging
└── tests/             # Test suite

Development commands

uv run ruff format   # code formatting
uv run ruff check    # linting
uv run ty check     # type checking
uv run pytest       # test suite

Repository

https://github.com/Alishahryar1/free-claude-code