Claude Code + OpenRouter = Free UNLIMITED Coding (No RAM Needed)

Overview

This guide explains how to run Claude Code via OpenRouter so you can use cloud-hosted models without downloading large local models or needing lots of RAM/VRAM. You authenticate Claude Code with a single OpenRouter API key and point the Anthropic/Claude endpoints to OpenRouter, which routes requests to free (or paid) cloud-hosted models.

What the video shows (high level)

• Use OpenRouter as a cloud backend for Claude Code so you can run Claude Code locally without heavy local models or high RAM/VRAM requirements.
• Authenticate Claude Code with a single OpenRouter API key and configure Anthropic endpoints to point to OpenRouter; OpenRouter will then route requests to available cloud models.

Step-by-step guide (core actions)

1. Install Claude Code (this guide assumes it is already installed).
2. Create a settings.json in a local .claude folder (or edit the global Claude config) and add the required variables:
   • OpenRouter API key
   • Anthropic base URL → set to the OpenRouter endpoint
   • Anthropic auth token → set to the OpenRouter key
   • Anthropic API key → set explicitly to blank
3. Create an OpenRouter account and generate an API key, then copy/paste it into settings.json.

Note: the presenter mentioned that $10 credit might be required in some cases, but they got it working without credit. If you encounter issues, try adding credit.

1. On OpenRouter’s models page, select a model and copy its model name into settings.json.
   • If a specific model is rate-limited or unavailable, use the general free models router OpenRouter/free, which routes requests among available free models and is more reliable.
2. Restart Claude Code or reload the session and confirm the model shows as OpenRouter/free (or your chosen model). Test prompts.

Example mapping (do not include real keys here):

anthropic:
  base_url: "https://api.openrouter.ai"        # OpenRouter endpoint
  auth_token: "openrouter_api_key_here"        # your OpenRouter key
  api_key: ""                                  # must be blank
model: "OpenRouter/free"                       # or a specific model name from OpenRouter

Practical notes, limitations, and tips

• This method avoids downloading heavy models — useful for Raspberry Pi, low-RAM laptops, or small VPSes.
• Free cloud models vary in capability and availability; many are rate-limited and performance is hit-or-miss.
• Local models (even large 20B–120B ones) still require significant VRAM and typically won’t match proprietary Sonnet/Opus 4.6 performance.
• You can switch models on the fly by replacing the model name in settings.json.
• OpenRouter also supports paid models (e.g., GPT 5.4 Pro) using the same configuration.

Models, performance, and cost analysis (mentioned)

• GPT-OSS — smallest OSS variant still needs roughly ~14 GB VRAM for local use.
• Qwen 3 Coder (480B) — extremely large; presenter quoted ~290 GB VRAM requirement, making it impractical locally.
• Step 3.5 Flash — highlighted as a good free/router model (in the top ~20% for coding/analysis; popular with Claude Code).
• MiniMax M2.5 — strongly recommended for coding by the presenter:
  • Ranked #1 for programming in their benchmarks.
  • Much cheaper than Opus for token costs:
    • MiniMax: cited at $0.30 per 1M input tokens, $1.20 per 1M output tokens.
    • Opus 4.6: cited at $5 per 1M input, $25 per 1M output.
• Sonnet 4.6 / Opus 4.6 — higher-performing commercial models; best-in-class quality but more expensive.

Troubleshooting tips

• If you hit rate limits, try a different free model or use the general router (OpenRouter/free).
• If you encounter authentication/errors:
  • Double-check the Anthropic fields in settings.json.
  • Ensure the Anthropic API key is blank and the auth token is set to your OpenRouter key.
• To apply the configuration globally, edit the global Claude settings.json (copy local block into the environment variables section).

Tools and platforms referenced

• Claude Code (Anthropic client/extension)
• OpenRouter (API/model router aggregator)
• Editor/terminal tools shown in the video: Cursor, terminal
• Alternatives mentioned: Ollama, self-hosted LLMs, running models on Raspberry Pi / VPS

Key takeaways

• You can run Claude Code “for free” by routing requests through OpenRouter without installing heavy local models.
• Expect variability in latency and availability; the free router endpoint is the most reliable option.
• MiniMax M2.5 is a good cost-performance choice for coding tasks if you want better-than-free-model results without Opus/Sonnet-level costs.

Main speaker / sources

• Video presenter / narrator (unnamed)
• Services/products referenced: Claude Code (Anthropic), OpenRouter
• Models mentioned: Qwen 3 Coder, GPT-OSS, Step 3.5 Flash, MiniMax M2.5, Sonnet 4.6, Opus 4.6, GPT 5.4 Pro