Free Claude Code: A Proxy That Turns $0 NVIDIA API Keys Into Unlimited Coding Agents

By Prahlad Menon 3 min read

Claude Code’s subscription model has a simple problem: you’re paying for infrastructure you might not need.

The Claude Code CLI and VS Code extension are excellent — fast, well-designed, deeply integrated with the Anthropic API. But they only work with Anthropic’s API. If you want to use a different model — say, NVIDIA’s free Nemotron 120B, or DeepSeek V4, or a local Ollama instance — you’re out of luck.

free-claude-code fixes this. It’s a proxy that sits between Claude Code and your model provider of choice. Claude Code sends its normal Anthropic API requests. The proxy translates them to whatever format your backend expects, forwards them, translates the response back, and returns it to Claude Code.

Claude Code never knows the difference.

What You Actually Get

This isn’t a toy project. With 37,000+ stars and 5,800+ forks, it’s become the de facto way to run Claude Code without Anthropic’s API:

17 backend providers:

  • Free tiers: NVIDIA NIM (40 req/min), OpenRouter (free models), Google Gemini, Mistral Experiment
  • Paid options: DeepSeek, Wafer, Kimi, Cerebras, Groq, Fireworks, Z.ai
  • Local: LM Studio, llama.cpp, Ollama

Per-model routing: Send Opus/Sonnet/Haiku requests to different providers. Run expensive tasks on DeepSeek V4 Pro, cheap tasks on local Ollama.

Native model picker: The proxy exposes a /v1/models endpoint. Claude Code’s /model command shows your configured models directly.

Codex support: Same proxy handles OpenAI’s Codex CLI and VS Code extension via the Responses API.

Remote access: Optional Discord and Telegram bots let you run Claude Code from your phone. Voice transcription via Whisper or NVIDIA NIM.

Admin UI: Local web interface at localhost:8082/admin for configuration. No config files to edit.

Quick Start: NVIDIA NIM Free Tier

The fastest path to free Claude Code:

1. Install

macOS/Linux:

curl -fsSL "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.sh?raw=1" | sh

Windows PowerShell:

irm "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.ps1?raw=1" | iex

This installs Claude Code and Codex if missing, then sets up the proxy.

2. Get NVIDIA NIM API Key

Go to build.nvidia.com/settings/api-keys. Sign up (free), generate a key.

Free tier limits: 40 requests/minute across all NIM models.

3. Start the Proxy

fcc-server

The terminal shows: Admin UI: http://127.0.0.1:8082/admin

4. Configure

Open the Admin UI. Paste your NVIDIA NIM key into NVIDIA_NIM_API_KEY. The default model is already set to nvidia_nim/nvidia/nemotron-3-super-120b-a12b.

Click Validate and Apply.

5. Run Claude Code

fcc-claude

That’s it. Claude Code runs through the proxy, talking to NVIDIA Nemotron 120B instead of Anthropic.

Available Models on NVIDIA NIM Free Tier

The free tier includes some impressive models:

ModelProviderNotes
nvidia_nim/nvidia/nemotron-3-super-120b-a12bNVIDIA120B MoE, default
nvidia_nim/moonshotai/kimi-k2.5MoonshotStrong reasoning
nvidia_nim/z-ai/glm5.1ZhipuChinese + English
nvidia_nim/minimaxai/minimax-m2.5MiniMaxFast inference

Browse all models at build.nvidia.com.

Per-Model Routing

The killer feature for cost optimization. In the Admin UI:

MODEL = nvidia_nim/nvidia/nemotron-3-super-120b-a12b  # fallback
MODEL_OPUS = deepseek/deepseek-v4-pro                 # heavy tasks
MODEL_SONNET = nvidia_nim/moonshotai/kimi-k2.5        # normal tasks
MODEL_HAIKU = ollama/llama3.2                          # quick tasks

When Claude Code requests claude-3-opus, the proxy routes to DeepSeek. When it requests claude-3-haiku, the proxy routes to local Ollama.

Same Claude Code UX. Different models per task weight.

Remote Access: Telegram Bot

The optional Telegram integration turns your phone into a Claude Code terminal:

  1. Create a Telegram bot via @BotFather
  2. Set TELEGRAM_BOT_TOKEN in the Admin UI
  3. Start the bot: fcc-telegram

Send a message to your bot. It forwards to Claude Code, returns the response. Voice messages get transcribed via Whisper or NVIDIA NIM before processing.

Discord works the same way with fcc-discord.

How It Compares

vs. Paying for Claude Code subscription: You’re trading Anthropic’s models (Claude 4/Opus/Sonnet) for free alternatives (Nemotron, Kimi, DeepSeek). If you don’t need Claude specifically — if you need “a good coding model” — this is pure savings.

vs. claude-code-local: claude-code-local is a native Anthropic API server for fully local inference on Apple Silicon. It’s faster (no translation layer) but requires a Mac with 96GB+ RAM for large models. free-claude-code is a proxy to external providers — works anywhere, no hardware requirements.

vs. OpenRouter directly: OpenRouter gives you one API key for many models, but Claude Code doesn’t speak OpenRouter’s format natively. free-claude-code handles the translation. OpenRouter is one of the 17 backends it supports.

vs. LiteLLM: LiteLLM is a general-purpose LLM proxy. free-claude-code is Claude Code-specific: it handles Anthropic’s Messages API, streams thinking tokens, manages tool calls, integrates with Claude Code’s /model picker. Purpose-built beats general-purpose.

The Tradeoff

You’re not getting Claude models. You’re getting whatever model your backend provides, accessed through Claude Code’s interface.

For many tasks — especially routine coding work — that’s fine. Nemotron 120B, Kimi K2.5, and DeepSeek V4 are all capable coding models. The Claude Code CLI is just a good terminal interface; the model behind it is swappable.

For tasks where you specifically need Claude’s reasoning or style, you’ll still want the real thing.

37,000 stars. 17 providers. One curl | sh. The Claude Code subscription model just got optional.