Video summary

I Gave CLAUDE CODE 1.6 Billion Free Tokens Every Day

Main summary

Key takeaways

Technology

Core claim / goal

  • The video argues that Anthropic / “Cloud code” is expensive because token quotas run out when you need them most, and the only official “fix” is paying more.
  • It claims a workaround: set up a local “Claude Code” environment that uses OmniRoute / OmniProxy to route requests to multiple providers’ free-tier models, enabling roughly ~1.6B free tokens/day (as claimed by the creator) without paying.

Tech setup & tools used (tutorial-style)

  1. Install OmniRoute

    • Website: omniroute.online
    • Install via a terminal command (requires Node.js first).
    • Launch locally on localhost.
    • Initial password: “CHANGE ME” (all caps), then change it.

  2. Configure OmniRoute providers

    • Uses OmniProxy → Providers with a large catalog:
      • 295 total providers
      • 126 free-tier providers (per the subtitle)
    • The creator emphasizes connecting the “best free providers” and notes the system will ignore models that fail compatibility checks.

  3. Add and validate multiple AI providers with free models

    • OpenRouter
      • Create an OpenRouter API key (via Google/GitHub login).
      • Set expiration short (creator suggests 1 day).
      • In OmniRoute, import only free models; validates by a “test model.”
      • Notes “OpenCode” models may require no authentication / are free.
    • Kiro AI
      • Authorized via Google login.
      • Tests Claude variants like Claude Sonnet 5 and Claude HiQ 4.5 (and others).
    • Anti-gravity
      • Authorized via Gmail; tested models include Claude Opus 4.6 thinking, Claude Sonnet 4.6, Gemini 3.6 Flash, etc.
    • Nvidia NIM
      • Create an Nvidia account and API key.
      • Import only free models.
      • Example mentioned: GLM 5.2 works (and some others may not).

Install & run “Claude Code” locally (with routing)

  1. Install Claude Code

    • Download command from the Claude Code site for Windows.
    • Wait 5–10 minutes depending on internet speed.

  2. Fix PATH / command not recognized

    • If Claude isn’t recognized, locate it in:
      • C:\Users\<username>\.local\bin\
    • Add that path to Windows environment variables → PATH.
    • Restart VS Code / the terminal and confirm Claude runs.

  3. Important execution order

    • Start OmniOut (local router client) before using Claude Code so the API key/config is active.
    • After restarting OmniRoute, re-launch OmniOut and continue setup.

Main configuration: create a “Combo” (model routing plan)

  1. Create first combo in OmniRoute

    • Go to “combos” and create a free stack (shown as $0).
    • Name it (example: “Amedia-AI”; the creator later uses the combo name in Claude Code config).

  2. Choose provider + model priority

    • Select a provider (example shown: Anti-gravity).
    • Set priority arrows / ordering (which models load first).
    • Option to “select model” to add specific models not in the initial provider list (examples like Opus 4.6 / Sonnet 4.6).

  3. Routing strategy

    • Default is priority.
    • The creator recommends round robin so it can reroute to the next best model automatically.

  4. Test the combo

    • A test run shows some models working and others erroring.
    • Guidance: ignore incompatible-model errors since OmniRoute will only use compatible ones.

  5. Create a combo API key

    • Generate an API key for the combo; store it safely since it won’t be shown again.

Connect Claude Code to the combo (final wiring step)

  1. In the Claude Code terminal

    • Paste a setup command.

  2. Critical edits before running

    • Replace the model name with the exact combo name from OmniRoute (no spaces).
    • Paste the OmniRoute / Combo API key into the command.

  3. Claude Code prompts setup questions

    • Use recommended settings.
    • Trust the folder.

  4. Validation

    • Asking “hi” works.
    • Asking which model it’s using returns the combo name (example: “Media AI” / “Amedia-AI”).
    • Checking OmniRoute’s control center shows successful requests and quota usage.

Demonstration / results claimed

  • Requests are shown as 100% successful in OmniRoute’s control center during the test.
  • Quota example shown: quota 90%, with unused quota for other models/providers until needed.
  • The system is described as switching providers/models automatically and quickly (milliseconds) when quota is exhausted.
  • A larger prompt is used to generate a website project:
    • Files are created by Cloud / Claude Code.
    • The website is started with npm run dev.
    • The creator says the generated site is “pretty decent” and can be tweaked.

Key sources/speakers (attribution)

  • Primary speaker: the YouTube video creator/host (no name provided in the subtitles).
  • Technical systems referenced: OmniRoute (OmniProxy), OpenRouter, Kiro AI, Anti-gravity, Nvidia NIM, and Claude Code / Anthropic Claude.

Original video