FCC is an open-source local proxy that lets Claude Code, Codex, and other agents run on free or cheap models instead of Anthropic's paid API.
What is Free Claude Code (FCC)? #
Free Claude Code, or FCC, is an open-source local proxy that sits between the Claude Code CLI and whatever model provider you actually want to pay for, or not pay for at all. Claude Code sends its requests thinking it’s talking to Anthropic’s API. FCC intercepts those requests, translates them into the format a different provider expects, and routes them wherever you’ve configured. The result is that you get the Claude Code interface and agentic workflow without needing an Anthropic subscription or API key at all.
TL;DR #
FCC is a local proxy, not a fork of Claude Code, that translates Anthropic-format API calls into formats used by other model providers, so Claude Code runs normally without ever touching Anthropic’s servers.- It supports roughly 25 providers, including Nvidia NIM, OpenRouter, Gemini, DeepSeek, Mistral, Grok, Cerebras, GitHub Models, Kimi, MiniMax, and GLM through z.ai, plus local inference through Ollama, LM Studio, and llama.cpp. - The project isn’t limited to Claude Code: it also works with Codex and Pi, meaning one proxy setup covers three coding agents. - A standout feature is per-tier model routing, where you can send Claude Code’s Opus-class requests to one model, Sonnet-class requests to another, and Haiku-class background tasks to a third, controlling both cost and reasoning effort independently. - Setup is handled through a browser-based admin UI at localhost, not config files, where you paste provider API keys, pick default models, and validate the connection. - FCC ships with Discord and Telegram bot integrations, including voice note transcription, so you can trigger coding tasks from your phone. - The project is MIT licensed and was reportedly sitting around 37,000 GitHub stars at the time it was covered, reflecting how much demand there is for alternatives to Anthropic’s usage caps.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Why does this exist? #
Claude Code is widely regarded as one of the strongest AI coding agents available, but Anthropic’s usage limits and pricing have become a real friction point. Weekly caps, an expensive Max tier, and users hitting limits mid-task have pushed a lot of developers, especially students and hobbyists, to look for ways to keep the Claude Code experience without the Anthropic bill. FCC solves this by decoupling the harness (the CLI, its tool-calling logic, its agentic loop) from the model actually doing the reasoning. You keep the interface you like and swap out the expensive backend for something free, cheap, or fully local.
How does the FCC proxy actually work? #
FCC runs as a local server, by default listening on port 8082. When you launch Claude Code through FCC’s wrapper command instead of the normal Claude Code launcher, all of its API traffic gets pointed at this local server instead of Anthropic’s endpoints. The proxy then handles the protocol translation between Anthropic’s “messages” format and the OpenAI-style format used by most other providers, including details like thinking blocks, tool calls, and token usage accounting. Because Claude Code’s tool-calling and agentic behavior depend on this translation being accurate, FCC has to replicate that plumbing closely enough that the agent behaves the same way it would against the real Anthropic API. From the user’s side, nothing about the workflow changes: the agent still reads files, edits code, runs commands, and reports progress the same way.
How do you install and set it up? #
Installation is a single command. On Mac or Linux, you run a curl command that pipes an install script from FCC’s repository into your shell. Windows users get an equivalent PowerShell one-liner. The installer checks for Claude Code, Codex, or Pi and installs them if they’re missing, and it sets up the proxy itself. For anyone wary of piping remote scripts into a shell, the install scripts are viewable in the repo before you run them, and the same command handles future updates.
Once installed, you get a few new commands, the main one being a server-start command that launches the proxy in your system tray on Windows and Mac, or in your terminal on Linux. Configuration happens entirely through a web-based admin UI in your browser rather than editing config files by hand. From there you paste in API keys for whichever providers you want to use, choose a default model, and validate the connection.
Requirements are minimal: a recent Node.js version (the project specifies Node 20.19, 22.13, or 24 and above) and API keys from whichever providers you plan to route requests to. Free options mentioned include Nvidia NIM’s free tier (which gives access to models like NeMoTron), OpenRouter’s free models, Gemini’s free API tier, GitHub Models, and the free tiers on Groq and Cerebras.
What can you do with per-tier model routing? #
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
One of FCC’s more interesting features is that Claude Code internally splits work across model tiers: Opus for heavy reasoning, Sonnet for general-purpose work, and Haiku for small background tasks. FCC lets you assign a completely different model to each tier. In practice, that means you could route Opus-tier requests to a large model like Kimi or GLM, keep Sonnet-tier work on something like a Gemini Flash model, and send Haiku-tier background tasks to a small, fast model on Groq, essentially for free. You can also tune reasoning effort per tier, cranking it up for the heavy model and turning it off for lightweight tasks. This lets you balance cost, speed, and quality without touching a single config file, all from the admin UI.
Does it work outside the terminal? #
Yes. Beyond the standard CLI usage, FCC can be pointed at from the Claude Code extension in VS Code by setting a couple of environment variables (an Anthropic base URL pointing at your local proxy, and an auth token). Similar integration exists for the Codex extension through its config file, and JetBrains editors are supported through ACP configuration. So the proxy isn’t limited to terminal use, it can sit behind whatever editor-integrated agent workflow you’re already using.
FCC also includes built-in Discord and Telegram bot support. After creating a bot through the Discord developer portal or Telegram’s BotFather, you paste the token into the admin UI’s messaging section, point it at a project directory, and you can then message your coding agent directly, including commands to check session status, stop a task, or clear context. Voice notes are supported too, with local Whisper transcription or transcription through Nvidia NIM, so a voice message like “fix that bug in the auth flow” can trigger the agent directly from a phone.
Is Free Claude Code worth using? #
For anyone frustrated by Claude Code’s usage caps or pricing, FCC is a genuinely well-built option. The combination of broad provider support, per-tier routing, a proper admin UI instead of manual config editing, editor integrations, and messaging bots in one MIT-licensed package is unusual for a project like this. It’s particularly appealing for students or hobbyists who like Claude Code’s agentic workflow but can’t justify paying for Anthropic’s Max plan. The tradeoff to understand clearly: you are not getting Anthropic’s actual Claude models for free. You’re getting the Claude Code harness, meaning its interface, tool-calling logic, and agent loop, running on top of other companies’ models. Quality will vary depending on which model you route requests to. Some limitations apply too: local models need enough context capacity to handle the agent’s system prompt and tool definitions, so very small local models may struggle, and not every model is equally good at tool calling, which is central to how well the agent performs. A few providers also have quirks (Cloudflare needs both an API token and account ID, for instance, and Codestral uses a separate key from the main Mistral key). Still, given how capable open and free-tier models have become, this is a reasonable tradeoff for a lot of people.
Frequently Asked Questions #
Is Free Claude Code (FCC) legal and does it violate Anthropic’s terms?
FCC doesn’t modify Anthropic’s software or bypass any paywall on Claude models themselves. It simply reroutes Claude Code’s API calls to other providers you have your own accounts and keys with. It’s open source under the MIT license.
Do I need an Anthropic account to use FCC?
No. Once FCC is configured with a provider API key and running as your proxy, you launch Claude Code through FCC’s command and it connects without ever logging into or authenticating with Anthropic.
Which models work best with FCC?
#
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
The transcript highlights larger models like Kimi and GLM as producing results close to genuine Claude Code performance for many tasks. Model choice matters most for tool-calling reliability, since that’s core to how agentic coding tools function.
Can I run FCC completely offline?
Yes, through local inference backends like Ollama, LM Studio, or llama.cpp. Performance will depend heavily on your hardware and the size of the local model, since the model needs enough context to handle the agent’s system prompt and tool definitions.
Does FCC only work with Claude Code?
No. It also supports Codex and Pi, so the same proxy setup can serve multiple coding agents rather than being limited to one tool.