Blog · How-to

How to Run Claude Code Free on Windows and Mac (OmniRoute Setup)

You're halfway through a project in Claude Code, and then: usage limit reached. It always hits right when you need it most, and Anthropic's answer is simple. Pay more, and go from Pro to Max. This guide skips that answer. With a free tool called OmniRoute, you can run Claude Code on free AI models, on your own computer, on Windows or Mac, with no VS Code and no credit card.

We set it up from scratch on a Mac on October 28, 2026, then gave it a real job. It built a full barber booking website in 1 minute 39 seconds, fixed its one bug in 37 seconds, and our OpenRouter bill for all of it was $0.00.

What OmniRoute is

OmniRoute is free, open source, has about eighty-two thousand stars on GitHub, and runs on your own computer. Claude Code makes one connection, to OmniRoute. Behind it, OmniRoute can reach more than three hundred AI providers, and over a hundred of them have free models.

It does three jobs for you:

  • One place for free providers. You add them once in a dashboard.
  • Smart routing. Each request goes to a model that's working right now.
  • Automatic fallback. When one model hits its limit, it moves on to the next one, so your work keeps going.

One correction before you start. You may have heard it gives you 1.6 billion free tokens a day. OmniRoute's own website says 1.51 billion a month. Still a lot, just not per day. (A token is a small chunk of text, roughly part of a word.)

Step 1: Install Node and OmniRoute

Install Node from nodejs.org. Pick the LTS version and click through the installer. Node is the free program OmniRoute runs on.

Now open your terminal. On a Mac, that's the Terminal app. On Windows, open PowerShell from the Start menu. Type npm install -g omniroute and press enter. You'll see some yellow warnings. That's normal, so ignore them. When it says "added", you're done.

Then type omniroute and press enter. It starts up and shows your dashboard address, localhost port 20128. Open that in your browser.

The paid button you don't need

On OmniRoute's website, the big button at the top says Try Cheaper Inference. That's a paid service from the same company, and it's why some people saw a price. You don't need it. The free version is the one you just installed, the self-hosted one, and you never enter a card.

Step 2: Add free providers

Log in with the default password, CHANGEME. Click Providers, search for OpenRouter, and click Add Connection. Make a free account at openrouter.ai, create a key, and paste it in.

Now the important switch: Import only free models. Turn that on, and paid models can't sneak in. Save it, switch the connection on, and import the models. Then press Test message, pick a model, and send. If it answers, it works.

Do the same for NVIDIA with a free key from build.nvidia.com. One thing we noticed while testing: some of NVIDIA's models in that list had already been switched off. Free models come and go, so always test one before you rely on it.

Step 3: Build your combo

A combo is a list of models that take turns. This is the part that keeps you going when a limit hits. Click Combos, then Create Combo, and name it free-stack. Add your models one by one. We added four:

  • NVIDIA's Nemotron, through OpenRouter
  • MiniMax, on NVIDIA
  • Poolside's Laguna, through OpenRouter
  • Nemotron again, straight from NVIDIA

For the strategy, pick Round Robin. Each request goes to the next model in line, and if one is busy or out of quota, OmniRoute skips it. Create the combo.

Last thing in the dashboard: open API Keys, create a key, call it claude-code, and copy it. That's how Claude Code will log in to OmniRoute.

Step 4: Install Claude Code

Back in the terminal. On a Mac, Claude Code installs with the one-line installer from the official setup docs. On Windows PowerShell, use the PowerShell line from the same page. When it says installed, type claude --version to check.

If Windows says "claude is not recognized", the install still worked. Windows just can't find it yet. Close PowerShell and open a new one, and that fixes it for most people. If it doesn't, the PDF has the one-line fix.

Step 5: The claude-free shortcut

Now you tell Claude Code to talk to OmniRoute instead of Anthropic. We made a shortcut called claude-free. It does four things:

  • Points Claude Code at OmniRoute on your own computer (localhost:20128).
  • Logs in with your OmniRoute key.
  • Picks your combo, free-stack, as the model.
  • Keeps its own settings folder, so your normal Claude Code stays exactly as it was.

On a Mac the shortcut lives in a settings file called .zshrc; on Windows it goes in your PowerShell profile. Both versions are in the free PDF, ready to paste. Then make a folder for your project and type claude-free. The first time, it asks a few setup questions. Pick a style, press enter, and when it asks if you trust this folder, choose yes.

Which model really answered?

Ask it: which model are you? It says it's Claude Code, powered by free-stack. That's because Claude Code tells every model that it's Claude. To see who really answered, open Request Logs in the dashboard. In our run it was MiniMax M3.1 Flash, on NVIDIA's free tier, and the cost was zero.

Up at the top, the app also says API Usage Billing. Don't worry about that. It just means it's using your OmniRoute key instead of a Claude subscription, and OmniRoute is billing you nothing.

Real job: a booking website

We asked it to build a booking website for a barber shop: the services and prices, pick a day and a time, type your name, then a confirmation screen.

It was done in 1 minute 39 seconds. And here's the part we liked. Halfway through, MiniMax hit its rate limit, and the combo quietly moved on to Poolside's Laguna, then to Nemotron, without stopping the build.

When we clicked through the page, everything worked except the very last step. Typing a name turned the Next button off instead of on. We told it that in plain English, and it fixed that one line in 37 seconds. After the fix, every step worked, from picking a haircut to the confirmation screen.

What we checkedResultVerdict
Full booking site from one prompt1:39Works
Model switches mid-buildMiniMax hit its limit, then Laguna, then NemotronWorks
First click-throughLast step broken (Next button logic flipped)Half
Bug fix in plain English37 secWorks
OpenRouter spend, before vs after$0.00Works

The catch: 3 things before you use it for everything

  1. Free models come and go. If something stops working, open the dashboard and press test again.
  2. Your messages go to whichever provider answers them, and some free tiers keep what you send. Use this for learning and your own projects, and keep client code and passwords out of it.
  3. These are other companies' models, not Claude. For everyday jobs they do fine. For your hardest work, the paid plan is still stronger.

One bonus tip. Out of the box, OmniRoute accepts requests from any device on your Wi-Fi. The PDF shows the one setting that keeps it private to your own computer.

Get the $0 Claude Code Card

So next time the limit hits halfway through your project, you don't have to stop. Type claude-free, and keep going. Every command for Mac and Windows is in one free PDF, so you can copy and paste instead of pausing the video. Comment COMBO on the video, or grab it below.

How we tested: Mac, OmniRoute 3.8.52, Claude Code 2.1.312, October 28, 2026. Combo of four free models from OpenRouter and NVIDIA, Round Robin. OpenRouter usage read from its API before and after: no change. The Windows commands come from the official docs and were not run on camera. One small test, not a study. Independent, not sponsored by or affiliated with OmniRoute or Anthropic.

Free PDF: The $0 Claude Code Card

The $0 Claude Code Card: every command from the video for Mac and Windows, the claude-free shortcut, the "not recognized" fix and the one setting that keeps OmniRoute private, so you can copy and paste instead of pausing.

Join 25,000+ readers of the Hyperautomation AI Report

Quick answers

Can I use Claude Code for free?

The Claude Code app itself installs for free, and with OmniRoute you can point it at free AI models instead of Anthropic's paid ones. In our test we built a full barber booking website and our OpenRouter account showed $0.00 spent. The catch: the answers come from other companies' free models, not Claude.

What is OmniRoute?

OmniRoute is a free, open-source tool with about eighty-two thousand stars on GitHub. It runs on your own computer and sits between Claude Code and more than three hundred AI providers, over a hundred of which have free models. When one model hits its limit, it moves on to the next one so your work keeps going.

Does OmniRoute give you 1.6 billion free tokens a day?

No. That number gets repeated a lot, but OmniRoute's own website says 1.51 billion free tokens a month, not a day. That's still a lot for learning and side projects, just not per day.

Why does OmniRoute's website ask me to pay?

The big button at the top of OmniRoute's website, Try Cheaper Inference, is a paid service from the same company. You don't need it. The free version is the self-hosted one you install with npm on your own computer, and you never enter a card.

How do I fix "claude is not recognized" on Windows?

The install worked; Windows just can't find Claude Code yet. Close PowerShell and open a new one, which fixes it for most people. If it still fails, add the install folder to your PATH with the one-line command in our free PDF, then open a new PowerShell.

Is it safe to use free AI models with Claude Code?

Your messages go to whichever provider answers them, and some free tiers keep what you send. Use it for learning and your own projects, and keep client code and passwords out of it. Also, out of the box OmniRoute accepts requests from any device on your Wi-Fi, so turn on the setting that keeps it private to your computer.