What it actually is
Free Claude Code is an Anthropic-compatible proxythat sits between Claude Code and the model. Claude Code thinks it's hitting Anthropic. The proxy quietly forwards the request to whichever provider you choose.
Tool calls, streaming, the /modelpicker, reasoning blocks — everything keeps working. You don't change how you use Claude Code. You just stop paying for it.
The honest version of the tradeoff
You're not getting Opus 4.7 for free. You're getting the best open source models — Kimi K2.6, GLM 5, DeepSeek V4. These get you about 90% of the way there for most coding tasks.
If you're already paying $200/month for Claude Max AND still hitting your limits every day, this is the best alternative on the market. Run the routine stuff on free models, save your Max plan for the truly hard problems.
Per-tier routing is built in, so MODEL_OPUS, MODEL_SONNET, and MODEL_HAIKU can each point at a different backend.
The full setup, step by step
You can do this. Follow each step in order. Don't skip ahead.
Step 1 — Install Claude Code itself
Open your terminal (Terminal on Mac, PowerShell on Windows) and paste this:
npm install -g @anthropic-ai/claude-code
If you don't have Node installed, grab it from nodejs.org first.
Step 2 — Install uv (a Python package manager)
The proxy is written in Python. uvhandles the install for you — you don't need to know Python.
On Mac or Linux, paste:
curl -LsSf https://astral.sh/uv/install.sh | sh uv self update uv python install 3.14
On Windows (PowerShell), paste:
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex" uv self update uv python install 3.14
Step 3 — Get your free NVIDIA NIM API key
Go to build.nvidia.com/settings/api-keys. Sign in (Google works), click Generate API Key, copy the key somewhere safe. You'll paste it in Step 6.
Step 4 — Install the proxy
Back in the terminal, paste:
uv tool install --force git+https://github.com/Alishahryar1/free-claude-code.git
Wait 30 seconds. It downloads and sets everything up automatically.
Step 5 — Start the proxy
Paste:
fcc-server
The terminal prints something like Admin UI: http://127.0.0.1:8082/admin. Cmd+click (Mac) or Ctrl+click (Windows) that link to open it in your browser.
Leave this terminal window running. Closing it shuts off the proxy.
Step 6 — Plug in your API key
In the admin page that just opened, find the NVIDIA_NIM_API_KEY field. Paste the key from Step 3. Click Validate, then click Apply. The default model is already set to GLM 4.7. You can swap it later.
Step 7 — Run Claude Code through the proxy
Open a new terminal window (leave the proxy running in the first one). Paste:
fcc-claude
This launches Claude Code, but every request now goes through your free proxy. You're done.
From now on, anytime you want to code with Claude for free, just run fcc-claude instead of claude.
Where to get free keys (pick ONE to start)
- NVIDIA NIM — free tier with GLM 4.7, GLM 5, Kimi K2.6, MiniMax M2.5. The easiest starting point.
- OpenRouter — free tier on certain models (look for
:freein the name). Pay-as-you-go on the rest. - Kimi (Moonshot) — go straight to the source for Kimi K2.6. Very cheap, not free.
- DeepSeek — direct API for DeepSeek V4. Dirt cheap, not free.
- Ollama / LM Studio / llama.cpp — fully local on your own machine. Free if you have a decent GPU.
You don't need all of these. NVIDIA NIM is the lowest-friction free option. Start there.
The /model picker still works
Inside Claude Code, type /modeland you can switch backends on the fly — same picker UX you already know. Run GLM 4.7 for routine work, switch to Kimi K2.6 when you want stronger reasoning. No restarts, no config edits.
When to skip this
- You're on Claude Max and not hitting your limits
- You need top-tier reasoning on every prompt — stay on Opus 4.7
- You hate managing local services — stay on Claude
- The tokens aren't your bottleneck — not your problem to solve
This is for the people running Claude Code 8 hours a day and hitting the Max wall by 2pm. That's where the math actually flips.
The bigger move
The real unlock isn't "free Claude Code." It's that Claude Code is becoming a protocol. The CLI is good enough that people are building tooling around it, not around a specific model.
Once you can swap the model behind the same UX, the model becomes a commodity and the agent harness becomes the moat. This proxy is the first popular tool to make that swap trivial.