Tools

Compose accepts OpenAI-style function tools on text models that support tool calling. The gateway passes tool definitions to the selected provider adapter and returns tool calls in OpenAI-compatible response shapes. Compose does not run the function. Your client or runtime owns permissions, sandboxing, side effects, and the follow-up request.

Chat completions example

curl https://api.compose.market/v1/chat/completions \
  -H "Authorization: Bearer $COMPOSE_MARKET_API_KEY" \
  -H "Content-Type: application/json" \
  -H "x-network-id: eip155:${evmNetworkId} | solana:${svmNetworkId}" \
  -d '{
    "model": "gpt-5.5",
    "messages": [
      { "role": "user", "content": "What is the weather in Rome?" }
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_weather",
          "description": "Get current weather for a city.",
          "parameters": {
            "type": "object",
            "properties": {
              "city": { "type": "string" }
            },
            "required": ["city"]
          }
        }
      }
    ],
    "tool_choice": "auto"
  }'

Tool calls return as:

{
  "tool_calls": [
    {
      "id": "call_...",
      "type": "function",
      "function": {
        "name": "get_weather",
        "arguments": "{\"city\":\"Rome\"}"
      }
    }
  ]
}

Send the tool result back as a tool message:

{
  "role": "tool",
  "tool_call_id": "call_...",
  "content": "{\"temperature_c\":18,\"condition\":\"clear\"}"
}

Responses example with two parallel calls

Responses tools are flat; they do not use Chat Completions’ nested function object:

{
  "model": "gpt-4.1",
  "input": "Compare the weather in Rome with the current time in Tokyo.",
  "parallel_tool_calls": true,
  "tools": [
    {
      "type": "function",
      "name": "get_weather",
      "description": "Get current weather for a city.",
      "parameters": {
        "type": "object",
        "properties": { "city": { "type": "string" } },
        "required": ["city"]
      }
    },
    {
      "type": "function",
      "name": "get_time",
      "description": "Get the current time in a city.",
      "parameters": {
        "type": "object",
        "properties": { "city": { "type": "string" } },
        "required": ["city"]
      }
    }
  ]
}

A model may return both calls in the same ordered response.output. Each call has two different stable identifiers:

[
  {
    "id": "fc_weather",
    "type": "function_call",
    "call_id": "call_weather",
    "name": "get_weather",
    "arguments": "{\"city\":\"Rome\"}"
  },
  {
    "id": "fc_time",
    "type": "function_call",
    "call_id": "call_time",
    "name": "get_time",
    "arguments": "{\"city\":\"Tokyo\"}"
  }
]

Execute both functions concurrently, then return both results together using the call_id values—not the output-item id values:

{
  "model": "gpt-4.1",
  "previous_response_id": "resp_...",
  "input": [
    {
      "type": "function_call_output",
      "call_id": "call_weather",
      "output": "{\"temperature_c\":18,\"condition\":\"clear\"}"
    },
    {
      "type": "function_call_output",
      "call_id": "call_time",
      "output": "{\"local_time\":\"09:30\"}"
    }
  ]
}

With store: false, replay the complete previous response.output first—including reasoning and both function-call items—then append both function_call_output items.

Provider behavior

The gateway normalizes the request shape. Provider behavior still differs:

Provider family	Notes
OpenAI	Chat requests with function tools and reasoning parameters may use the Responses wire internally when required by the model.
Azure OpenAI	Uses the same OpenAI-style message lowering path for tools, tool choice, response format, and custom params.
Gemini	Tool-call state can include provider metadata needed for later turns; Compose preserves it where the adapter exposes it.
OpenAI-compatible providers	Tool support depends on the selected model and provider. Check the catalog row before relying on it.

External clients

OpenCode and similar IDEs send large tool payloads. The external gateway keeps those fields intact and passes AI SDK wrapper fields through the core normalizer where they are supported. External content-filter errors are sanitized in the response body. Diagnostic headers still show the selected provider and model when safe:

x-upstream-provider: azure
x-upstream-model: gpt-5.5
x-public-model: gpt-5.5

Overview

External Use

Families

Modalities

Guides

Chat completions example

Responses example with two parallel calls

Provider behavior

External clients

​Chat completions example

​Responses example with two parallel calls

​Provider behavior

​External clients

​Related

Chat completions example

Responses example with two parallel calls

Provider behavior

External clients

Related