Tool / Function Calling

PromptGate routes tool / function-calling requests transparently. You send the OpenAI-shaped tools definition once, PromptGate translates it for whichever provider answers, and you get a uniform tool_calls array back — same shape no matter who served it.

Send a tool call

POST /api/{projectUuid}/{endpointSlug}
Authorization: Bearer pg_live_…
Content-Type: application/json

{
  "messages": [{ "role": "user", "content": "What's the weather in Berlin?" }],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Look up current weather for a city.",
      "parameters": {
        "type": "object",
        "properties": {
          "location": { "type": "string" },
          "unit": { "type": "string", "enum": ["c", "f"] }
        },
        "required": ["location"]
      }
    }
  }],
  "tool_choice": "auto"
}

tool_choice accepts the OpenAI vocabulary: "auto", "none", "required", or an explicit {"type":"function","function":{"name":"…"}}. Anthropic and other backends are translated transparently.

Normalized response

When the model decides to call a tool, the response carries a flat tool_calls array:

{
  "ok": true,
  "data": {
    "id": "chatcmpl-…",
    "content": "",
    "model": "gpt-4o-mini",
    "finish_reason": "tool_calls",
    "tool_calls": [
      {
        "id": "call_a1",
        "name": "get_weather",
        "arguments": { "location": "Berlin", "unit": "c" }
      }
    ],
    "usage": { "prompt_tokens": 12, "completion_tokens": 5, "total_tokens": 17 }
  }
}

The arguments field is already parsed into a real object. OpenAI returns it as a JSON string upstream; PromptGate parses it for you. Anthropic returns it as an object; nothing to parse.

tool_calls is null when the model produced regular text. Use that as the branch condition in your client.

Provider translation

Provider	Inbound shape	Translation
OpenAI / Mistral / Groq / Together / Ollama	OpenAI-shaped	Pass-through
Anthropic	OpenAI-shaped (`{ type: "function", function: { … } }`)	Translated to `{ name, description, input_schema }`. `tool_choice` mapped: `"auto"` → `{type:"auto"}`, `"required"` → `{type:"any"}`, named → `{type:"tool", name}`
Cohere	(TBD)	Pass-through if already Cohere-shaped

Already-native shapes pass through untouched, so if you target a single provider you can also send its raw format.

Observability

Live Logs row shows a TOOL · N chip when tool_calls is non-empty.
Detail modal has a dedicated Tool Calls tab listing each call with name, id, and pretty-printed arguments.
gateway_logs.response_body.tool_calls is queryable (the same JSON column already used for plain responses).

Schema validation

If an endpoint has output_schema configured and the response carries tool calls, PromptGate skips schema validation (tool calls aren’t text — they’re structured invocations). Plain-text responses on the same endpoint still validate as before.

Streaming

v1 ships with non-streaming tool calls only. Streaming responses pass through, but the SSE chunks aren’t yet re-mapped to a normalized tool-call delta protocol. Streaming + tool calls is a follow-up.

Limitations

Wrapper API (/v1/chat/completions) — the AI Wrapper currently passes tools through but doesn’t normalize tool_calls in the OpenAI-compatible response shape yet (OpenAI clients will see them in their native shape, which is fine for OpenAI-targeted apps).
Cohere v2 doesn’t speak the OpenAI tools shape; its native tools API is not yet bridged.
Tool result messages (role tool with tool_call_id) are forwarded as-is. PromptGate doesn’t inspect or transform them — your client/agent is responsible for the round-trip.

© Akyros Labs LLC. All rights reserved.