额度又见底了？这个开源项目，给 AI 编程助手接上 61 家"加油站"
用 Claude Code 写代码的朋友，多少都体验过这种崩溃。
改代码改到关键时刻，屏幕上突然弹出一句"额度已用完"。思路断了，活儿停了，只能眼睁睁看着进度条干着急。
最近在 GitHub 上刷到一个开源项目——free-claude-code（简称 FCC），专治这个病。它在本地架了一层"模型调度中枢"，把 61 家模型服务商和 12 个主流编程智能体接到一起。官方的口号相当直白：每月 13 亿+ 免费 token 可用。
它到底是个啥？
一句话：FCC 是一个跑在你自己电脑上的"中间层"。
左边接你的编程工具（Claude Code、Codex、Cline、Aider……），右边接各家模型服务商（NVIDIA NIM、OpenRouter、Groq、Gemini、DeepSeek、Kimi、通义……），中间由一个本地 Admin 网页统一调度。
配好之后，在 Claude Code 里输入 /model，就能看到一整排可用模型，随手切换。
先说清楚：它是独立开源项目（AGPL-3.0 协议），和 Anthropic 官方没有任何关系，也不是破解工具——它只做调度，走的都是各家官方 API 和订阅，项目方还特别强调遵守服务商条款。
12 个编程智能体，共用一个模型池
FCC 目前支持 12 个 coding agent：Claude Code、Codex、Pi、OpenCode、Cline、Hermes、DeepSeek Harness、Grok Build、Muse Code、Aider、Antigravity CLI、VS Code Chat。
每个都有专属启动命令：fcc-claude、fcc-codex、fcc-aider……装一次，全家桶共享同一个模型池，哪个顺手用哪个。
61 家服务商，怎么选？
README 里给了完整目录，我帮你归成四类：
- 白嫖党：NVIDIA NIM（官方教程第一步就配它）、OpenRouter 免费档、Groq 等，注册就送额度；
- 订阅党：ChatGPT 订阅、GitHub Copilot、Kimi Code、通义灵码编程计划……把手头已有的订阅用满；
- 云 API 党：Gemini、DeepSeek、智谱 GLM、Azure、Bedrock……按量付费，稳定优先；
- 本地党：Ollama、LM Studio、llama.cpp，模型跑在自己显卡上，一分钱不花，代码不出门。
四步上手
- 一行命令安装，Windows / macOS / Linux 都有官方脚本；
- 启动 FCC：开始菜单点开图标，或终端运行 fcc-server；
- 去 build.nvidia.com 申请一个免费 API key，粘进 Admin 界面；
- 终端敲 fcc-claude，开干。
全程不用改任何工具的配置文件，对新手相当友好。
真正的杀手锏：永不断粮
免费额度谁都有，FCC 的价值在调度能力上：
- 故障自动切换：当前模型报错或限流，FCC 自动切到你预设的下一个模型，对话不中断，不用重来；
- 四档分层路由：Fable / Opus / Sonnet / Haiku 四个档位可以分别指定不同模型——重活给最强模型，杂活交给免费模型；
- 省 token 外挂：可选安装 RTK 和 Headroom，官方说法是最高可省 90% 的 token 用量；
- 浏览器原生会话：不用开终端，在浏览器里选个文件夹就能直接跑 Codex。
手机上也能遥控写代码？
配一个 Discord 或 Telegram 机器人，聊天软件秒变遥控器：出门在外发条消息，让 agent 继续干活；发 /stats 看进度，/stop 随时叫停。
甚至支持语音：直接发一段语音，本地 Whisper 或 NVIDIA 接口转成文字，agent 收到就开工。
上车前，三件事必须知道
- 隐私：你的代码会发送给你配置的服务商。免费额度某种意义上是"用数据换的"，核心商业项目建议走付费 API 或本地模型；
- 合规：免费额度和限流由各服务商控制，随时可能调整，别把关键流程押在单一免费源上；
- 别迷信：再稳的调度也救不了含糊的需求描述，用清楚的话给 AI 派活，这个基本功还是你自己的。
写在最后
AI 编程工具的竞争，正在从"谁的模型最强"变成"谁的调度最稳、成本最低"。FCC 这类项目的思路很清醒：不押注任何一家，把选择权拿回自己手里。
额度焦虑治好了，写代码的快乐回来了。你的 AI 编程额度一个月烧多少？
注：本文信息基于该 GitHub 项目 2026 年 10 月的 README，免费额度、支持的服务商和工具列表以项目最新文档为准；该项目与 Anthropic 官方无关，请遵守各服务商使用条款。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
quota is back? This open source project, connect AI programming assistant to 61 gas stations.
Friends who write code with Claude Code have experienced this kind of collapse.
Change the code to a critical moment and suddenly a line "quota is out of service" appears on the screen. I've lost my mind, I've stopped my work, and I can't wait to see it.
Recently, an open source project, free-claude-code (FCC), was painted on GitHub to treat the disease. It built a local model movement hub, bringing 61 model service providers together with 12 mainstream programming AI agent. The official slogan is quite straightforward: 1.3 billion + free token per month.
What is it?
One sentence: FCC is a middle layer running on your own computer.
To the left, pick up your programming tools (Claude Code, Codex, Cline, Aider...) and to the right, the model service providers (NVIDIA NIM, OpenRouter, Groq, Gemini, DeepSeek, Kimi, Qwen...) with a local Admin page in the middle.
When you're ready, enter /Model in Claude Code, you can see a whole row of usable models, hand-to-hand.
Let's be clear: it's an independent open source project (AGPL-3.0 protocol), and it has nothing to do with Anthropic official, nor is it a cracking tool -- It is only a dispatch, all of which go to official APIs and subscriptions, and project parties also place special emphasis on compliance with service provider provisions.
12 programming AI agent, one model pool
FCC currently supports 12 parties: Claude Code, Codex, Pi, OpenCode, Cline, Hermes, DeepSeek Harness, Grok Build, Muse Code, Aider, Antigravity CLI, VS Code Chat.
Each has an exclusive start-up order: fcc-claude, fcc-codex, fcc-aider once, the whole family can share the same model pool and which one to use.
Sixty-one service providers, what's the choice?
README gave you a full catalogue, and I'll put you in four categories:
- free access Party: NVIDIA NIM (with the official curriculum first), OpenRouter free tier, Groq et al., registered to quota;
- Subscribe parties: ChatGPT subscriptions, GitHub Copilot, Kimi Code, Qwen elixir programming program.
- The Yun API Party: Gemini, DeepSeek, Zhipu GM, Azure, Bedrock... at paid, priority for stabilization;
- Local party: Ollama, LM Studio, llama.cpp, model running on its own GPU, with no money, no code.
Four steps.
- An official script is available in Windows / MacOS / Linux;
- Start FCC: start the menu point icon or run the fcc-server;
- Go to bild.nvidia.com and apply for a free API key, stick to the Admin interface;
- Fcc-claude, dry.
beginner is quite friendly without changing the configuration of any tool.
The real killer: Always food.
free quota has everybody. FCC values in dispatch capability:
- Automatic breakover: the current model is wrong or rate limit, the FCC automatically cuts to the next model that you pre-set, the conversation is uninterrupted and does not have to be repeated;
- Four-storey routers: Fable / Opus / Sonnet / Haiku, each of which can assign different models - reliving to the strongest model and handing over the chores to the free model;
- Province token hang: optional installation of RTK and Headroom, which is officially up to 90 per cent of token usage;
- Browser native session: without opening the terminal, select a folder in the browser to run directly to Codex.
You can write code on your phone?
With a Discod or Telegram robot, chat software seconds to remote control: send out a message outside so that angent can continue working; send /stats to see progress, /stop to stop at any time.
It even supports voice: direct voice, local Whisper or NVIDIA interface to text, and start work when you receive it.
Three things must be known before you get in the car.
- Privacy: Your code will be sent to your configured server. free quota is in a sense "for data" and the core business project proposes a paid API or a local model;
- Compliance: free quota and rate limit are controlled by various service providers and may be adjusted at any time to avoid placing critical processes on a single free source;
- Don't be superstitious: you can't save a vague description of needs and give AI life in clear words.
At the end.
The competition for AI programming tools is moving from "who's strongest model" to "who's most stable and least costly." The FCC has a clear idea of such projects: not bet on any family, take the choice back to yourself.
quota's anxiety is cured, and the happy writing is back. How much does your AI program quota burn a month?
Note: This information is based on a list of README, free quota, supported service providers and tools for October 2026 for the GitHub project, which is not official for Anthropic and is based on the latest project files; please comply with the service provider usage terms.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.