ByteCrypt

Claude Code with OpenRouter Custom Models

2 min read

Claude Code is not locked to Anthropic. The CLI happily talks to any endpoint that speaks the Anthropic Messages protocol, so you can route it through OpenRouter and run whatever model you want — including non-Claude models, or in my case a stealth model (stealth/ox-alpha) that I also use as my Hermes brain.

How it works

Claude Code reads its config from ~/.claude/settings.json. The whole trick is four environment variables under the env key:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://openrouter.ai/api",
    "ANTHROPIC_AUTH_TOKEN": "<your openrouter key>",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_MODEL": "stealth/ox-alpha",
    "ANTHROPIC_SMALL_FAST_MODEL": "stealth/ox-alpha",
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "200000"
  }
}

What each line does:

  • ANTHROPIC_BASE_URL — instead of https://api.anthropic.com, every request goes to OpenRouter. Note the URL is openrouter.ai/api (not /api/v1) because OpenRouter exposes an Anthropic-compatible skin at /api/v1/messages, and Claude Code appends /v1/messages itself.
  • ANTHROPIC_AUTH_TOKEN — the OpenRouter API key, sent as a Bearer token. ANTHROPIC_API_KEY must stay empty so Claude Code doesn’t try to use it or trigger OAuth login.
  • ANTHROPIC_MODEL — the default model for everything interactive.
  • ANTHROPIC_SMALL_FAST_MODEL — the background model Claude Code uses internally (session titles, autocompletion-ish tasks). I set it to the same model so everything routes to one place.
  • CLAUDE_CODE_MAX_CONTEXT_TOKENS — optional but useful: since the model name is unknown to Claude Code, it can’t look up the context window and will nag about auto-compact. This tells it to assume 200k tokens.

Getting there

  1. Install: npm install -g @anthropic-ai/claude-code. On Fedora with npm’s script-blocking, allow the postinstall first: npm config set allow-scripts=@anthropic-ai/claude-code --location=user — otherwise the native binary never downloads and claude errors out.
  2. Keep the API key out of the settings file if the file might get shared; mine is sourced from ~/.secrets (export or="sk-or-v1-...") and injected by a small script.
  3. Verify: claude -p "Reply with exactly: OK" --model stealth/ox-alpha should print just OK.

Caveats

  • Claude Code prints [claude-code:unrecognized_model] warnings for custom models — harmless. It just doesn’t know the model’s context window, hence the CLAUDE_CODE_MAX_CONTEXT_TOKENS shim above.
  • Switching models later is just editing ANTHROPIC_MODEL to any OpenRouter slug (deepseek/deepseek-chat, google/gemini-3.6-flash, etc.) — no proxy needed, OpenRouter translates on their side.
  • You pay OpenRouter per-token for whatever model you pick, so this trades Anthropic billing for OpenRouter credits. Free-tier models exist too (:free suffix) with rate limits.

Graph View