Tool Calling

The model proposes a tool name and JSON arguments. Your server, not the model, validates permissions and performs the database, network, or write operation.

Declare a tool

# Start with a read-only tool to verify model compatibility safely.
curl https://sprelaytoken.com/v1/chat/completions \
  -H "Authorization: Bearer $SPRELAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "messages": [
      {"role": "user", "content": "What is the weather in London today?"}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_weather",
          "description": "Get weather for a city",
          "parameters": {
            "type": "object",
            "properties": {
              "city": {"type": "string", "description": "City name"}
            },
            "required": ["city"],
            "additionalProperties": false
          }
        }
      }
    ]
  }'

Field	Meaning
tools	Tools the model may request during this call
type: function	Declares a function-style tool
name	Stable name used to select your handler
description	Helps the model decide when to use it
parameters	JSON Schema for allowed arguments
required	Arguments that must be present
additionalProperties: false	Rejects undeclared arguments

JSON cannot contain comments, so preserve its punctuation when editing.

Execution loop

1. Send messages and tool definitions.
2. Check whether the model returned a tool call.
3. Validate the name against an allowlist and the arguments against the Schema.
4. Apply business authorization and a timeout.
5. Execute the approved server-side handler.
6. Send the result back with the original tool-call ID.
7. Read the model's final answer.

Treat all model-generated arguments as untrusted input. Never concatenate them directly into SQL, shell commands, or unrestricted URLs. Require confirmation or idempotency for writes, cap tool-call loops, and avoid returning credentials or internal stack traces.