Use Tool Calling with Routor

Tool calling works exactly like OpenAI’s function-calling API - the same tools array, the same tool_calls shape in the response. What’s different is what happens underneath: as soon as a request includes tools, Routor filters to tool-capable models and reranks them by measured function-calling accuracy (BFCL - Berkeley Function Calling Leaderboard) for the specific kind of tool use in that request, not just by price.

Basic Request

const response = await client.chat.completions.create({
  model: "auto",
  messages: [{ role: "user", content: "What's the weather in Tokyo?" }],
  tools: [
    {
      type: "function",
      function: {
        name: "get_weather",
        description: "Get the current weather for a location",
        parameters: {
          type: "object",
          properties: { location: { type: "string" } },
          required: ["location"],
        },
      },
    },
  ],
});

const call = response.choices[0].message.tool_calls?.[0];
if (call) {
  console.log(call.function.name, call.function.arguments);
}

This is unmodified OpenAI tool-calling shape - if you already have tool-calling code against OpenAI, pointing it at Routor requires no changes beyond the base URL and key.

What Happens Differently

The instant a request has a non-empty tools array, Routor:

Filters the candidate chain to tool-capable models only (same as it filters to vision-capable models when an image is attached)
Classifies the kind of tool use - a single expected call, a web-search-shaped tool, a memory/vector-store tool, several tools that might fire in parallel, or a multi-turn tool conversation already in progress - and reranks the remaining candidates by that scenario’s BFCL score specifically, not a generic overall score
Still applies the normal difficulty scoring on top - a one-tool weather lookup and a multi-step agentic tool chain don’t land on the same tier just because both have tools set

The practical effect: two requests with the same tool definitions but different complexity can route to different models, and the model chosen is one that’s actually measured to be good at tool calls in that shape, not just the cheapest model that technically supports function calling.

Requiring a Minimum Tool-Calling Quality

If your use case can’t tolerate a weak tool-caller (an agent loop where a malformed call breaks the next step, for example), set a floor directly in the request:

{
  "model": "auto",
  "messages": [...],
  "tools": [...],
  "routor_bfcl_min": 0.7
}

Models below that BFCL score for the request’s scenario are dropped from the candidate chain entirely - see the API reference for the exact field name and behavior if it’s changed since this was written. The same constraint can be set permanently on a routing profile instead of every request, so an API key always enforces it.

Debugging a Tool-Routing Decision

POST /v1/routing/debug accepts the same tools array and returns the scenario it detected and the BFCL score of the model it would pick, without spending a real request:

curl https://api.routor.io/v1/routing/debug \
  -H "Authorization: Bearer sk-routor-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Search the web for the current Bitcoin price",
    "tools": [{"type":"function","function":{"name":"web_search","description":"Search the web","parameters":{"type":"object","properties":{"query":{"type":"string"}}}}}]
  }'

Where to Go Next

Full request/response reference - every routing-related body field
Build your first app - the non-tool starting point this guide builds on
Use with Claude Code - Claude Code’s own tool calls (file edits, bash, etc.) go through this same path

Getting Started

Core Concepts

Dashboard

Playground

Providers

Guides

Use Tool Calling with Routor

Basic Request

What Happens Differently

Requiring a Minimum Tool-Calling Quality

Debugging a Tool-Routing Decision

Where to Go Next

​Basic Request

​What Happens Differently

​Requiring a Minimum Tool-Calling Quality

​Debugging a Tool-Routing Decision

​Where to Go Next

Basic Request

What Happens Differently

Requiring a Minimum Tool-Calling Quality

Debugging a Tool-Routing Decision

Where to Go Next