model-router

An OpenAI-compatible proxy that classifies each conversation turn by complexity and routes it to the most cost-effective Claude model. Sits between your AI gateway (e.g. OpenClaw) and Anthropic's API.

No dashboard, no database, no framework, no telemetry. Just classify, route, log.

Before you start

• OpenClaw: This proxy works with OpenClaw 2026.3.2+. Install OpenClaw first if you don't have it.
• Anthropic API key: Sign up at console.anthropic.com and create an API key.
• System access: You'll need sudo to install the systemd service (or run it manually).

How it works

Your app / OpenClaw (model: router/auto)
        ↓
  localhost:3456/v1/chat/completions
        ↓
  [Classify last user message]
  [Pick model: Haiku / Sonnet / Opus]
        ↓
  Anthropic API (/v1/messages)
        ↓
  [Translate response → OpenAI format]
        ↓
  Return to caller

The proxy accepts requests in OpenAI format, classifies the last user message, translates the request to Anthropic format, forwards it, and translates the response back.

Classification

Classification uses a three-tier approach:

1. Rule-based fast path (instant, no API call)

These fire immediately without touching an LLM:

Rule	Tier	Model
/opus in message	complex	Opus
/cheap in message	simple	Haiku
router-test-simple / router-test-standard / router-test-complex	simple / standard / complex	Haiku / Sonnet / Opus
Heartbeat / [System Message]	standard	Sonnet

2. Haiku LLM classification (~200-500ms)

For everything else, the proxy sends just the user message to Claude Haiku with a short system prompt and asks it to reply with one word: simple, standard, or complex. This gives much better accuracy than keyword matching — Haiku understands intent, not just keywords.

The classification prompt defines the tiers as:

• simple: greetings, short factual questions, trivial lookups, yes/no questions, small typo fixes
• standard: normal coding tasks, bug fixes, moderate features, explanations, code review, single-file changes, tool use
• complex: architecture design, large multi-file refactors, multi-system integration, deep reasoning, long-form writing, formal proofs

Cost is negligible (~$0.0002 per classification). Latency is 200-500ms, unnoticeable against the main model call.

3. Keyword fallback

If the Haiku classification call fails (network error, rate limit, etc.), the classifier falls back to keyword matching: reasoning keywords → complex, complexity keywords → complex, tool-use keywords → standard, short messages → simple, default → standard.

Logs show which path was taken: (test hook), (haiku: complex), or (keyword fallback: default).

Format translation

The proxy handles full bidirectional translation between OpenAI and Anthropic formats:

• System messages: extracted from the messages array → Anthropic top-level system field
• Tool calls: OpenAI tool_calls ↔ Anthropic tool_use content blocks
• Tool results: OpenAI role: "tool" messages ↔ Anthropic tool_result content blocks (grouped into user messages)
• Tool definitions: OpenAI function.parameters ↔ Anthropic input_schema
• Stop reasons: end_turn → stop, max_tokens → length, tool_use → tool_calls
• Usage: input_tokens → prompt_tokens, output_tokens → completion_tokens

Installation

Prerequisites

• Node.js 18+
• An Anthropic API key

Build

git clone https://github.com/hopthings/model-router.git
cd model-router
npm install
npm run build

Run

ANTHROPIC_API_KEY=sk-ant-... npm start

The proxy starts on http://127.0.0.1:3456 (localhost only).

Configuration

Edit config.json in the project root. Changes take effect on next request (no restart needed).

{
  "port": 3456,
  "tiers": {
    "simple": "claude-haiku-4-5-20251001",
    "standard": "claude-sonnet-4-20250514",
    "complex": "claude-opus-4-6"
  },
  "rules": {
    "shortCharsSimple": 80,
    "forceStandardIfLikelyToolUse": true
  },
  "overrides": {
    "forceComplexRegex": "^/opus\\b",
    "forceSimpleRegex": "^/cheap\\b"
  }
}

Field	Description
port	Port the proxy listens on
tiers.simple	Anthropic model ID for simple messages
tiers.standard	Anthropic model ID for standard messages
tiers.complex	Anthropic model ID for complex messages
rules.shortCharsSimple	Messages shorter than this (chars) route to simple
rules.forceStandardIfLikelyToolUse	Floor tool-use messages to standard tier
overrides.forceComplexRegex	Regex to force complex tier
overrides.forceSimpleRegex	Regex to force simple tier

The tiers values are Anthropic model IDs (not prefixed with anthropic/).

OpenClaw integration

Add the proxy as a provider in your OpenClaw config:

{
  "models": {
    "providers": {
      "router": {
        "baseUrl": "http://127.0.0.1:3456/v1",
        "api": "openai-completions",
        "apiKey": "local-no-auth",
        "models": [{"id": "auto", "name": "auto"}]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "router/auto",
        "fallbacks": ["anthropic/claude-sonnet-4-20250514"]
      }
    }
  }
}

Running as a systemd service

Create /etc/systemd/system/model-router.service:

[Unit]
Description=Model Router Proxy
After=network.target

[Service]
Type=simple
User=ubuntu
Environment=ANTHROPIC_API_KEY=sk-ant-...
WorkingDirectory=/path/to/model-router
ExecStart=/usr/bin/node dist/server.js
Restart=always
RestartSec=3

[Install]
WantedBy=multi-user.target

sudo systemctl daemon-reload
sudo systemctl enable model-router
sudo systemctl start model-router

# Check logs
journalctl -u model-router -f

Testing

Health check

curl http://localhost:3456/health
# → {"status":"ok"}

Non-streaming

curl -s http://localhost:3456/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"hi"}],"max_tokens":50}' | jq .

Test routing

# Should route to Haiku
curl -s http://localhost:3456/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"router-test-simple"}],"max_tokens":50}' | jq .

# Should route to Opus
curl -s http://localhost:3456/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"router-test-complex"}],"max_tokens":50}' | jq .

Streaming

curl -N http://localhost:3456/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"hi"}],"max_tokens":50,"stream":true}'

Logging

Every request logs one line to stdout:

[router] 13:32:01 "router-test-simple" → simple → claude-haiku-4-5-20251001 (test hook)
[router] 13:32:15 "Write an article about AI governance" → complex → claude-opus-4-6 (haiku: complex)
[router] 13:32:30 "hi" → simple → claude-haiku-4-5-20251001 (haiku: simple)
[router] 13:33:00 "Add Stripe payment integration" → complex → claude-opus-4-6 (haiku: complex)

When running as a systemd service, view logs with journalctl -u model-router.

Project structure

model-router/
├── src/
│   ├── server.ts          # HTTP server, single POST endpoint
│   ├── classifier.ts      # Haiku LLM classifier + rule-based fast path + keyword fallback
│   ├── translator.ts      # OpenAI ↔ Anthropic format translation
│   ├── anthropic.ts       # Anthropic API client
│   ├── config.ts          # Config loader
│   └── types.ts           # Shared type definitions
├── config.json            # Routing config (user-editable)
├── package.json
└── tsconfig.json

License

MIT