free-claude-code (often shortened to FCC) is one of the busiest repos on GitHub trending this week, with about 59,000 stars and 9,300 forks at the time of writing. The pitch is catchy, but the name oversells it: it does not hand out free Claude. It is a local proxy that lets Claude Code and ten other coding agents talk to models from other providers, many of them on free tiers. This explainer covers how it works, what you can and cannot expect, and the terms and data risks you should weigh before installing it. It is not a guide to abusing anyone's quota.
TL;DR
| Question | Answer |
|---|---|
| What is it? | A local proxy and admin UI that makes other providers' models look like the API your agent client expects |
| Does it provide Claude models? | No. The README says it is independent and not endorsed by Anthropic |
| Is it free? | The software is AGPL-3.0-only; the repo claims "1.3B+ free tokens every month" across providers, but limits are set by each provider |
| Which clients? | The README lists 11: Claude Code, Codex, Pi, OpenCode, Cline, Hermes, DeepSeek Harness, Grok Build, Muse Code, Aider and VS Code Chat |
| Where does my code go? | To the provider you pick, with your key, through a proxy on your machine |
| Biggest risk | Free-tier data handling and terms you did not read, plus running an installer from the internet |
What it actually does
Claude Code speaks Anthropic's Messages API. If you change the base URL it calls, it will send the same requests somewhere else. FCC runs a server on your machine (fcc-server, or a tray app on Windows and macOS) that accepts those requests and translates them into whatever API the chosen provider offers, then translates the streamed answer back. Launchers such as fcc-claude, fcc-codex and fcc-opencode start each client already pointed at the proxy.
From the README, the main features are:
- A catalog of 61 providers, including NVIDIA NIM, OpenRouter, Groq, Together AI, DeepInfra, xAI, Alibaba Cloud and QwenCloud plans, GitHub Copilot, Amazon Bedrock, Vertex AI, Azure OpenAI, Ollama and custom endpoints.
- Fallback models: after retries fail it tries the next configured model. The README warns that a failed request "may reach and consume usage from more than one provider before succeeding", which matters for privacy as well as quota.
- A model picker exposed inside Claude Code's
/modelcommand, and routing of the Opus, Sonnet and Haiku slots to different backing models. - Optional token savers, RTK and Headroom, claimed to cut token use "up to 90%".
- Browser sessions, Discord and Telegram front ends, and voice notes through local Whisper or NVIDIA NIM transcription.
The default setup in the README uses an NVIDIA NIM key and the model nvidia_nim/nvidia/nemotron-3-super-120b-a12b. In other words, your "Claude Code" session is driven by the Claude Code harness but answered by a different model.
Terms of service: what is and is not claimed
The README's headline says the providers are "ToS-friendly" and that FCC "follows provider terms and removes integrations if they stop being allowed." It also adds notes such as: Kimi Code and QwenCloud Coding Plan keys are for personal interactive coding-agent use under those providers' own guidelines.
Read those carefully. They are claims by the maintainer, not guarantees:
- Per-provider terms differ. A plan sold for personal interactive use may not allow routing through a shared proxy, scripts, or other people's sessions. The README's own notes show the restrictions exist.
- Anthropic's terms are a separate question. FCC drives Anthropic's Claude Code client but, per the README, not Anthropic's models or credentials (Anthropic models need a separately billed API key), so the usual concern is the client license and any acceptable-use language in it. Read the Claude Code legal terms yourself rather than relying on a README.
- Do not use other people's accounts. Tools that pool or rotate keys are the classic route to bans. FCC supports multiple numbered keys per provider; use only keys you own and are allowed to use.
If the provider ever changes its free tier or terms, FCC's behavior can change overnight. The README itself says "free-tier availability and limits are controlled by each provider and may change."
Data risks, stated plainly
- Your code leaves your machine. Everything in the agent's context, including files it reads and command output, is sent to the provider. Free tiers commonly allow prompt logging and sometimes training; some prohibit sensitive data. Look up each provider's data policy before use.
- Fallbacks widen exposure. With a fallback list, one failed request can be replayed to a second or third provider.
- The proxy holds your keys. FCC stores provider API keys in its admin config. The README offers Proxy Authentication (a bearer token) for the local server; turn it on, and do not expose the port.
- The installer is curl-to-shell. The documented command is
curl -fsSL ... | sh. The README itself links the install scripts for review; read them first, or clone the repo and install from source. - Supply chain. A fast-moving repo with optional integrations (RTK, Headroom, Discord, Telegram) has a large dependency surface. Pin a version and review updates.
- Weaker models, real changes. A smaller free model driving an agent that can edit files and run commands makes mistakes. If you wouldn't give it an unreviewed shell, don't give it yours. A tool like AgentBeam from the explainx.ai team is built to stop agents before they take dangerous actions.
How it compares
| Option | Models | Cost | Data path | Fit |
|---|---|---|---|---|
| Claude Code with Anthropic | Claude | Plan or API pricing | Anthropic | Best fidelity; see our commands reference |
| FCC with a free tier | Whatever the provider offers | Free within limits | Third-party provider | Experiments, throwaway repos |
| FCC with local Ollama | Local weights | Hardware only | Stays on your machine | Private work, weaker models |
| Native open-source agents | Any | Varies | Your choice | See OpenCode 2.0 |
If your goal is cheaper tokens rather than free ones, cut waste first. Our posts on ColGrep cutting Claude Code token use and Sonnet 5.5 cache pricing show cheaper levers that keep your data with a vendor you've vetted.
What people are asking
Is the 1.3B free tokens figure real? It is the maintainer's aggregate across providers' free tiers. Each provider sets its own limit, so your personal number depends on which keys you hold and what they allow today.
Will it work as well as real Claude? Not likely. Claude Code's prompts and tool schemas are tuned for Claude. Other models may follow tool formats less reliably, and quality varies a lot. The README stresses preserving interleaved thinking and tool use, but the model on the other end still decides.
Why AGPL? The AGPL-3.0-only license means that if you modify FCC and let others use it over a network, you must share your changes. That suits a community project but may be a blocker for companies, so check with legal before adopting it internally.
Is this safe for a company laptop? Treat it like any unvetted developer tool: get approval, avoid credentials and customer data, and ask whether your policy allows sending code to the chosen provider.
Why it went viral, and what it signals
Three things explain the star count. First, the cost of agentic coding has become a daily irritation: heavy Claude Code users hit plan limits, and a repo promising a way around them lands well. Second, the market for hosted open-weight models has matured, with several providers now offering free or cheap tiers for models good enough to run an agent loop. Third, the client harness has become a product in its own right; the README's list of eleven interchangeable agents reflects a world where the interface and the model are chosen separately.
For builders this is the useful takeaway. If a proxy this simple can swap the model under a mature agent client, then your own tooling should assume the model is replaceable. Keep prompts, skills and tool definitions portable, test them against more than one model, and measure the real cost per completed task rather than the sticker price per token. A free model that needs three retries and a human review is not free.
For vendors, the lesson cuts the other way: harness lock-in is weak, so pricing and terms for the client software matter. Expect providers to keep tightening acceptable-use language around personal versus shared use, which is exactly the wording the README keeps flagging in its provider notes.
Try it safely
If you still want to experiment: use a throwaway repository with no secrets, create a dedicated API key with a low quota, enable proxy authentication, bind to localhost only, read scripts/install.sh first, and pick a provider whose free-tier data terms you have read. For the broader skill of working with agents, see our loop engineering guide and the MCP primer.
Bottom line
FCC is a useful piece of plumbing that shows how portable the agent client layer has become: the harness and the model are now separate choices. That portability is the real story. The free tokens come with other parties' terms, and the safety of your code depends on choices FCC leaves to you. Treat it as an experiment tool, not as a free Claude.
Related reading
- Claude Code commands: the complete reference
- OpenCode 2.0 upgrade guide
- ColGrep: a 2B local code search agent that cuts Claude Code tokens
- Sonnet 5.5 cache read price cut
- Loop engineering for coding agents
- What is MCP
- Official: free-claude-code on GitHub, Claude Code docs
Star counts, features and provider lists reflect the repository README on October 13, 2026 and change quickly.