Claude Code Router: Run Claude Code on Any Model

​

Prerequisites

• Node.js ≥ 20.12 (npx works without a global install)
• Claude Code installed
• An Opper account — free to create; the login step below opens sign-up if you don’t have one

​

Quick start

npx @opperai/cli login    # one-time OAuth sign-in in your browser
npx @opperai/cli launch claude

​

Switch the model

opper launch --model openai/gpt-5.4 claude
opper launch --model alibaba:global/kimi-k2.7-code claude

opper launch claude --resume

​

Manual configuration (without the CLI)

export ANTHROPIC_BASE_URL=https://api.opper.ai/v3/compat
export ANTHROPIC_AUTH_TOKEN=your-api-key
export ANTHROPIC_DEFAULT_OPUS_MODEL=anthropic/claude-opus-5
export ANTHROPIC_DEFAULT_SONNET_MODEL=anthropic/claude-sonnet-5
export ANTHROPIC_DEFAULT_HAIKU_MODEL=gemini/gemini-3.5-flash
claude

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.opper.ai/v3/compat",
    "ANTHROPIC_AUTH_TOKEN": "your-api-key",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "anthropic/claude-sonnet-5"
  }
}

​

Show gateway models in the /model picker

export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1

​

Claude Agent SDK and CI

env:
  ANTHROPIC_BASE_URL: https://api.opper.ai/v3/compat
  ANTHROPIC_AUTH_TOKEN: ${{ secrets.OPPER_API_KEY }}

​

Keep inference in the EU

​

Coming from OpenRouter or claude-code-router?

• From OpenRouter — same mechanism, different base URL: set ANTHROPIC_BASE_URL=https://api.opper.ai/v3/compat and pass your Opper key as ANTHROPIC_AUTH_TOKEN. You gain EU-hosted routing, inference at provider rates, and per-call tracing when you turn it on (Control Plane).
• From claude-code-router — the open-source proxy runs locally and needs its own config file and process. With Opper the router runs server-side: the two env vars above are the entire setup, and they work anywhere Claude Code runs, including CI.

​

What changes compared to Anthropic direct

• Billing. Routed sessions are billed as API usage through Opper at provider rates. A Claude Pro or Max subscription does not apply while a gateway credential is active.
• Model IDs. The opus / sonnet / haiku aliases resolve through the ANTHROPIC_DEFAULT_* mapping; for everything else use full IDs from the catalog.
• Feature parity. Some Anthropic-endpoint conveniences (prompt caching among them) depend on the route serving your request. Anthropic documents the gateway mechanism but does not officially support non-Claude models in Claude Code, so expect polish to vary by model — frontier models and strong open-weight tool-callers work well in practice.
• Prompt caching on non-Claude models. Claude Code starts the system prompt with an attribution block that carries its version and a fingerprint of the prompt. Providers that cache by prompt prefix, such as the open-weight models, read that block first, so conversations can’t share the cached system prompt. Before Claude Code v2.1.181 the block also changed on every request, so those versions got almost no cache hits on these models. Set CLAUDE_CODE_ATTRIBUTION_HEADER=0 to leave the block out (see Claude Code’s environment variables).
• Reliability. Set a dynamic route (dynamic/<name>) as the model in --model or an ANTHROPIC_DEFAULT_* variable to get a fallback chain (Control Plane): if the primary model fails or a provider has an outage, the call falls through to the next model in the route.

​

Troubleshooting

• “Model not found” — use the full model ID exactly as listed in the catalog; aliases only resolve through the ANTHROPIC_DEFAULT_* mapping.
• Auth errors — run opper whoami; re-run opper login if the session expired. If Claude Code was previously signed in to an Anthropic account, /logout once and check /status shows the Opper base URL.
• Context length errors — context windows differ per model; check the catalog and pick a model sized for your session.
• Checking spend — every routed session is metered with per-call cost, latency, and token accounting under Logs in your Opper dashboard.