Run Claude Code for Free: The Open-Source Proxy That Routes It to Free Models

There is now a free, open-source way to keep the Claude Code experience you like and stop paying Anthropic for every token. It works by quietly swapping the brain behind the interface, and more than 34,000 developers have already starred the project on GitHub.

Here’s the thing. Claude Code is one of the best terminal coding agents around, but the bill is real. An Anthropic subscription starts at a monthly fee, and if you run on the API, heavy agent loops can burn through credits fast. So a developer named Ali Shahryar built something clever: a small proxy that sits between Claude Code and the model, and lets you point that interface at free or local models instead.

The project is called Free Claude Code. Let’s break down exactly what it does, what “free” really means here, and how to set it up in about ten minutes.

What Free Claude Code Actually Is

Claude Code talks to Anthropic using a specific protocol called the Messages API. Free Claude Code is a drop-in proxy that speaks that same protocol. Your Claude Code client thinks it is talking to Anthropic. In reality, the proxy is forwarding your requests to whichever model you chose, then translating the responses back into the exact shape Claude Code expects.

That one trick unlocks a lot. The interface, the keyboard flow, the agent behaviour, the slash commands, the model picker, all of it stays the same. Only the engine underneath changes.

One thing worth being honest about up front: this does not give you Anthropic’s Claude models for free. You are keeping the Claude Code tool and routing it to other models, mostly open-weight ones from providers that offer a free tier. If you specifically want Opus or Sonnet quality, this is not a loophole for that. What it is, is a way to get a capable coding agent running at zero cost.

The project supports 17 different model backends, including NVIDIA NIM, OpenRouter, Google AI Studio, DeepSeek, Mistral, Groq, Cerebras, Z.ai, Kimi, and local runners like Ollama, LM Studio, and llama.cpp. You pick one, paste a key, and go.

Why Developers Are Paying Attention

Two things happened at roughly the same time, and together they made this idea suddenly practical.

First, the open-weight models got genuinely good at coding. Models like MiniMax M2.7, DeepSeek, GLM, and Kimi are now strong enough on real coding tasks that routing your agent to them is no longer a downgrade you suffer through. They are not Opus, but for a large slice of everyday work, the gap is smaller than you’d expect.

Second, the free inference tiers got generous. NVIDIA NIM now hands out free API access to dozens of hosted models with no credit card required, and OpenRouter keeps a rotating collection of free models. Put those together with a proxy that speaks Claude Code’s language, and you have a coding agent that costs nothing to run.

That combination is why the repo crossed 34,000 stars and 5,000 forks. Developers had been looking for exactly this.

What You’ll Need Before You Start

Nothing exotic. Here is the short list:

• Claude Code already installed. The proxy launches your existing claude command, so you need it on your machine first.
• A terminal you are comfortable with. macOS, Linux, or Windows PowerShell all work.
• One free API key from a provider like NVIDIA NIM or OpenRouter. This takes about two minutes and needs no payment details.
• Optional: a decent GPU if you want to run models entirely on your own machine through Ollama instead of a cloud provider.

Step-by-Step: Set Up Free Claude Code

The setup has four stages. Install the proxy, start it, configure a provider in the admin panel, then launch Claude Code through it. Take them in order.

Step 1: Install the proxy

Open your terminal and run the one-line installer for your system.

On macOS or Linux:

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

On Windows (PowerShell):

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

It is good practice to read any install script before piping it into your shell. Both files are public in the repo, so open them once and skim before you run.

Step 2: Start the proxy server

Once the install finishes, start the local server:

fcc-server

It prints the address it is running on and an admin URL, usually http://127.0.0.1:8082/admin. Keep this terminal window open. The proxy needs to stay running while you work.

Step 3: Open the Admin UI and add your provider key

Open that admin URL in your browser. This is a local-only control panel, reachable only from your own machine, where you manage everything. No config files to edit by hand.

1. Pick a provider. For a quick start, NVIDIA NIM is the easiest free option.
2. Grab a free key (see the next section for exactly where).
3. Paste it into the matching field, for example NVIDIA_NIM_API_KEY.
4. Set the MODEL field to a model you want, then click Validate and Apply.

Step 4: Launch Claude Code through the proxy

In a new terminal, run:

fcc-claude

This reads your proxy settings, points Claude Code at the local server, and launches the normal claude experience. You are now coding through a free model. To confirm it worked, open the /model picker inside Claude Code and you should see the models you configured.

If you use VS Code or a JetBrains IDE instead of the terminal, the repo has a short config snippet for each. Same idea, you just set a few environment variables so the editor extension points at http://localhost:8082 instead of Anthropic.

Choosing Your Model: Cloud or Local

You have two real paths here, and the right one depends on whether you have a capable GPU.

Option A: Cloud models (free API key)

This is the simplest route and works on any machine. You generate a free key from a provider and let them do the heavy compute.

• NVIDIA NIM: Sign up for the free NVIDIA Developer Program, open any model in the catalog, and generate a key. You get free access with a rate limit of around 40 requests per minute and no credit card. The catalog includes coding-capable models like DeepSeek, Kimi, GLM, MiniMax, and Qwen.
• OpenRouter: Create a key and point MODEL at one of its free models. Handy when you want to test several models behind a single key.

Option B: Local models (your own hardware)

If you have a GPU with enough memory, you can run the model fully offline through Ollama. Nothing leaves your machine, there are no rate limits, and there is no API key at all.

1. Install Ollama and pull a model, for example: ollama pull llama3.1
2. Start it: ollama serve
3. In the Admin UI, set MODEL to the same tag, prefixed with ollama/, like ollama/llama3.1.

Local is the most private and the most genuinely free option. The trade-off is that model quality is capped by what your hardware can run, and a 7B or 8B model will not match a large cloud model on complex tasks.

A Quick Model and Provider Cheat Sheet

The Honest Caveats

This is a clever tool, but it is not magic. A few things you should know before you rely on it.

The models are not Claude. Open-weight models have closed a lot of the gap on coding, but on hard reasoning, large refactors, and tricky multi-file tasks, you will still feel a difference versus Opus or Sonnet. Match the model to the job.

Free tiers have limits and can change. NVIDIA’s free tier carries a rate limit and is intended for development and testing, not production traffic. Free models can also be deprecated with little notice, so build your workflow to be model-flexible. If one model disappears, you should be able to swap in another from the Admin UI in seconds.

It is a community project, not an Anthropic product. Free Claude Code is independent and open-source under an MIT license. That is a strength for transparency, but it also means support is community-driven and things move fast. Check the repo’s issues page if something breaks.

Think about where your code goes. When you use a cloud provider, your prompts and code travel to that provider’s servers. If you are working with sensitive code, the local Ollama route keeps everything on your machine.

Who This Is Actually For

If you are a heavy Claude Code user on a paid plan and you mostly do straightforward work, you probably do not need to switch your whole workflow. Your subscription is doing its job.

But this is a strong fit if you are a student or hobbyist who wants a capable coding agent without a monthly bill, a developer in a region where subscriptions are expensive relative to local income, someone who wants to experiment with open models inside a familiar interface, or a privacy-focused coder who wants everything running locally through Ollama.

It is also just a great way to learn how these tools actually work under the hood. Once you see Claude Code happily running on a free DeepSeek or GLM model, the whole stack stops feeling like a black box.

The Real Takeaway

The interesting shift here is not really about saving money, although you will. It is that the interface and the model have come unbundled. The agent you code in and the brain that powers it no longer have to come from the same company. Free Claude Code is a tidy proof of that idea, and it takes about ten minutes to try.

Grab the project at github.com/Alishahryar1/free-claude-code, pick a free provider, and point your terminal somewhere new.

While you are optimizing your setup, it is worth pairing this with a way to see what your agent is actually doing. Our guide to CodeBurn, the free dashboard that shows exactly where your Claude Code tokens go, pairs nicely with this once you start running heavier sessions.

The interface you love and the model you pay for don’t have to be the same thing anymore. One small proxy, and Claude Code runs free.

Sources and Further Reading

1. Free Claude Code on GitHub (Ali Shahryar)
2. NVIDIA NIM model catalog (build.nvidia.com)
3. NVIDIA: free NIM access for Developer Program members
4. OpenRouter free models collection
5. Ollama (run models locally)
6. Guide to free AI models on NVIDIA NIM (rate limits and caveats)