router-control-public is a small LLM routing proxy for chat workloads.

It focuses on one job: classify a request, choose a profile and model, resolve a provider, and forward the call through an OpenAI-compatible API surface.

• Local rule-based request classification
• Profile-driven model selection
• Capability-aware provider filtering
• Model pool indirection per provider
• Fallback routing with lightweight failure tracking
• OpenAI-compatible POST /v1/chat/completions

• GET /health
• GET /providers
• GET /config/all
• GET /config/profiles
• GET /config/profiles/:name
• POST /config/profile
• POST /classify
• POST /v1/chat/completions

npm install
cp .env.example .env
npm run build
npm start

The server listens on ROUTER_PORT and defaults to 3000.

Set provider credentials with environment variables such as:

• OPENAI_API_KEY
• ANTHROPIC_API_KEY
• GOOGLE_API_KEY
• DEEPSEEK_API_KEY
• GROQ_API_KEY
• GITHUB_MODELS_TOKEN
• OPENROUTER_API_KEY
• OLLAMA_BASE_URL
• OMLX_BASE_URL

You can also use OPENAI_API_KEYS or ANTHROPIC_API_KEYS as comma-separated key pools.

Routing hints can be passed with x-router-* headers:

• x-router-feature
• x-router-agent-id
• x-router-session-id
• x-router-parent-agent-id
• x-router-project
• x-router-model-override

Legacy x-openclaw-* aliases are still accepted for compatibility.

curl http://127.0.0.1:3000/v1/chat/completions \
  -H 'content-type: application/json' \
  -H 'x-router-feature: conversation' \
  -d '{
    "messages": [
      { "role": "user", "content": "Explain how fallback routing works." }
    ]
  }'

Routing behavior is defined in:

• src/config/router-config.yaml
• src/config/profiles.yaml
• src/config/models.yaml
• src/config/model-pools.yaml

npm test

Architecture and runtime notes are in ARCHITECTURE.md and docs/architecture/router-architecture.md.