How to run Claude Code, Codex and 9 other coding agents for free (FULL GUIDE) A GitHub repo with 61.8k stars You keep the Claude Code interface, but the expensive model behind it gets swapped for free ones 61 providers, 1.3B+ free tokens per month, 11 coding agents on one model catalog Repo: github.com/Alishahryar1/free… Here’s how to install it in 5 minutes 1. How it works FCC is a local proxy. Instead of going to Anthropic, Claude Code talks to a server on your machine • Your agent sends a request to localhost:8082 • FCC forwards it to the provider and model you picked • The response comes back in the format Claude Code expects Claude Code doesn't notice anything changed. Even the /model picker shows your FCC models 2. What you get • 61 providers in one UI: NVIDIA NIM, OpenRouter, Groq, DeepSeek, Gemini, Ollama, LM Studio and more • 11 agents on one catalog: Claude Code, Codex, Pi, OpenCode, Cline, Hermes, Aider and others • Per-tier routing: strong model for Opus, cheap one for Sonnet, local one for Haiku • Auto fallback: if a model dies, FCC switches to the next one mid-turn • Control from terminal, VS Code, JetBrains, Discord or Telegram, even by voice notes 3. Setup in 4 steps > Step 1. Install Open a terminal and paste ONE line. Copy the whole thing, don't change anything, press Enter • Mac: the Terminal app (Cmd+Space, type Terminal) • Linux: any terminal • Windows: PowerShell (press Win, type PowerShell) Mac / Linux: curl -fsSL raw.githubusercontent.com/Al… | sh Windows: & ([scriptblock]::Create((irm raw.githubusercontent.com/Al…))) That's it. It downloads and installs everything by itself, no links to open When it asks which coding agents you want, pick at least one, for example Claude Code > Step 2. Start the server • Windows / macOS: open the "Free Claude Code" app from your desktop or Start menu • Linux: run fcc-server and keep the terminal open The Admin UI opens automatically in your browser > Step 3. Add a free provider • Get a free key at build.nvidia.com/settings/ap… • In Admin → Providers → NVIDIA NIM → Add key → Save • Default model is nvidia_nim/nvidia/nemotron-3-super-120b-a12b, or pick any other in the dropdown • Click Apply > Step 4. Launch your agent • fcc-claude for Claude Code • fcc-codex for Codex • fcc-opencode, fcc-pi, fcc-cline, fcc-aider work the same way The launchers start the server and set up the environment for you. Something broke? Run fcc-doctor 4. The setup that actually saves money Don't route everything to one model. Go to Admin → Model Config • Opus tier → your strongest model • Sonnet tier → a cheaper hosted one • Haiku tier → a local model via Ollama or LM Studio • Add Fallback Models in order, so rate limits never stop your session Most coding tasks (boilerplate, renames, tests, explaining code) don't need a frontier model Save the expensive one for hard problems 5. Where it works best Prototyping, refactors, CRUD, side projects, anything high-volume For hard architecture work, a weak model will still struggle, so switch to a stronger one there Like/RT/Bookmark if this was helpful!

How to run Claude Code, Codex and 9 other coding agents for free (FULL GUIDE) A GitHub repo with 61.8k stars You keep the Claude Code interface, but the expensive model behind it gets swapped for free ones 61 providers, 1.3B+ free tokens per month, 11 coding agents on one model catalog Repo: github.com/Alishahryar1/free… Here’s how to install it in 5 minutes 1. How it works FCC is a local proxy. Instead of going to Anthropic, Claude Code talks to a server on your machine • Your agent sends a request to localhost:8082 • FCC forwards it to the provider and model you picked • The response comes back in the format Claude Code expects Claude Code doesn't notice anything changed. Even the /model picker shows your FCC models 2. What you get • 61 providers in one UI: NVIDIA NIM, OpenRouter, Groq, DeepSeek, Gemini, Ollama, LM Studio and more • 11 agents on one catalog: Claude Code, Codex, Pi, OpenCode, Cline, Hermes, Aider and others • Per-tier routing: strong model for Opus, cheap one for Sonnet, local one for Haiku • Auto fallback: if a model dies, FCC switches to the next one mid-turn • Control from terminal, VS Code, JetBrains, Discord or Telegram, even by voice notes 3. Setup in 4 steps > Step 1. Install Open a terminal and paste ONE line. Copy the whole thing, don't change anything, press Enter • Mac: the Terminal app (Cmd+Space, type Terminal) • Linux: any terminal • Windows: PowerShell (press Win, type PowerShell) Mac / Linux: curl -fsSL raw.githubusercontent.com/Al… | sh Windows: & ([scriptblock]::Create((irm raw.githubusercontent.com/Al…))) That's it. It downloads and installs everything by itself, no links to open When it asks which coding agents you want, pick at least one, for example Claude Code > Step 2. Start the server • Windows / macOS: open the "Free Claude Code" app from your desktop or Start menu • Linux: run fcc-server and keep the terminal open The Admin UI opens automatically in your browser > Step 3. Add a free provider • Get a free key at build.nvidia.com/settings/ap… • In Admin → Providers → NVIDIA NIM → Add key → Save • Default model is nvidia_nim/nvidia/nemotron-3-super-120b-a12b, or pick any other in the dropdown • Click Apply > Step 4. Launch your agent • fcc-claude for Claude Code • fcc-codex for Codex • fcc-opencode, fcc-pi, fcc-cline, fcc-aider work the same way The launchers start the server and set up the environment for you. Something broke? Run fcc-doctor 4. The setup that actually saves money Don't route everything to one model. Go to Admin → Model Config • Opus tier → your strongest model • Sonnet tier → a cheaper hosted one • Haiku tier → a local model via Ollama or LM Studio • Add Fallback Models in order, so rate limits never stop your session Most coding tasks (boilerplate, renames, tests, explaining code) don't need a frontier model Save the expensive one for hard problems 5. Where it works best Prototyping, refactors, CRUD, side projects, anything high-volume For hard architecture work, a weak model will still struggle, so switch to a stronger one there Like/RT/Bookmark if this was helpful!

Every time someone says "ai can't write good code", I know that they can't write good code.

to me it was always bonkers insane that in order to connect service A with service B i need to ... paste an api key ??? how come as a fucking industry we didnt invent a better way to connect deployments with services

turns out deepseek is not infinite too when you are working on 1348139848913 things like a madman I WANT INFINITE COMPUTE PLS

BREAKING: Grok Bot’s App Store downloads surged 728.4% last month. Grok Bot is winning people over worldwide. Phenomenal growth. 🔥🚀

ok but what the fuck is this model name

I’m starting to think the biggest advantage of Opus 5.5 isn’t intelligence. It’s how much work you can get done before hitting the usage limit wall.

most days i'm bouncing between a handful of agent tools (claude code, cowork, codex, cursor, grok bot) giving them different projects to see how they handle things and how the model capabilities are evolving