Tools

AI Decisions Gateway

ai_decisions_gateway MCP

Classify text with typed questions and get each answer's confidence

Answer typed questions about a text with OpenAI's Decisions API, which picks answers instead of writing them: faster and cheaper than ai_gateway for classifying, routing, flagging or scoring, and every answer comes with how sure the model is. Use ai_gateway when something has to be written. The text to judge goes in user_prompt. The questions go in structured_output_json_schema, one per field, with the field's description as its question, and system_prompt holds framing shared by every question. A field name cannot contain a dot. Put a description on every field: a field without one is asked by its name. Each field must be one of: boolean (yes/no); enum of strings or whole numbers (pick one, up to 255 options); array of an enum (yes/no per option); number with minimum 0 and maximum 1 (the probability itself); integer levels 0..N with a description per level in anyOf/const (a score, up to 10 levels); or an object of these. At most 200 questions per text, counting one per option of an array field. structured_output mirrors the schema, with each field replaced by {"value": ..., "confidence": ..., "probabilities": ..., "score": ...}, so it does not validate against the schema itself. confidence runs from 0 to 1; a probability field has none, being a probability already. A score's value is its continuous score rounded to the nearest level, which can land between two likely levels on one the model gave no weight to: read probabilities for the likeliest level.

Playground

Log in Log in to try this tool. The request uses your session.

Credits

Fixed cost per call plus input tokens only -- no output tokens. Prefeed agents are billed on top, at their own rates.

These usage SKUs can be charged on a call.

Schemas

item is what you send, config how the tool behaves, and the response is what you get back.

{
  "prefeed": [],
  "user_prompt": "My parcel never arrived and nobody answers. Refund me now!"
}

{
  "prefeed": [],
  "system_prompt": "",
  "structured_output_json_schema": {
    "properties": {
      "team": {
        "description": "Which team should handle this ticket?",
        "enum": [
          "billing",
          "logistics",
          "tech"
        ]
      },
      "angry": {
        "description": "Is the customer angry?",
        "type": "boolean"
      },
      "topics": {
        "description": "Which topics does the ticket raise?",
        "items": {
          "enum": [
            "refund",
            "delay",
            "damage"
          ]
        },
        "type": "array"
      },
      "refund_likelihood": {
        "description": "How likely is the customer to ask for a refund?",
        "maximum": 1,
        "minimum": 0,
        "type": "number"
      },
      "severity": {
        "anyOf": [
          {
            "const": 0,
            "description": "cosmetic"
          },
          {
            "const": 1,
            "description": "annoying"
          },
          {
            "const": 2,
            "description": "blocking"
          },
          {
            "const": 3,
            "description": "data loss"
          }
        ],
        "description": "How severe is the issue?",
        "type": "integer"
      }
    },
    "required": [
      "team",
      "angry",
      "topics",
      "refund_likelihood",
      "severity"
    ],
    "type": "object"
  },
  "model_settings": {
    "boolean_threshold": 0.5,
    "model": "gpt-6-luna"
  }
}

{
  "structured_output": {}
}

Endpoints

See running a long job in the background for the polling flow.

cURL

curl -X POST "http://agents.botify.com/{organization}/{project}/ai_decisions_gateway/process" \
  -H "Authorization: Bearer $BOTIFY_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
  "item": {
    "user_prompt": "My parcel never arrived and nobody answers. Refund me now!"
  },
  "config": {
    "structured_output_json_schema": {
      "properties": {
        "team": {
          "description": "Which team should handle this ticket?",
          "enum": [
            "billing",
            "logistics",
            "tech"
          ]
        },
        "angry": {
          "description": "Is the customer angry?",
          "type": "boolean"
        },
        "topics": {
          "description": "Which topics does the ticket raise?",
          "items": {
            "enum": [
              "refund",
              "delay",
              "damage"
            ]
          },
          "type": "array"
        },
        "refund_likelihood": {
          "description": "How likely is the customer to ask for a refund?",
          "maximum": 1,
          "minimum": 0,
          "type": "number"
        },
        "severity": {
          "anyOf": [
            {
              "const": 0,
              "description": "cosmetic"
            },
            {
              "const": 1,
              "description": "annoying"
            },
            {
              "const": 2,
              "description": "blocking"
            },
            {
              "const": 3,
              "description": "data loss"
            }
          ],
          "description": "How severe is the issue?",
          "type": "integer"
        }
      },
      "required": [
        "team",
        "angry",
        "topics",
        "refund_likelihood",
        "severity"
      ],
      "type": "object"
    }
  }
}'

Over MCP

This tool is exposed as the MCP tool ai_decisions_gateway. An MCP client passes organization, project and item as tool arguments. Set up MCP →

Raw metadata

• GET /agents/ai_decisions_gateway — this page as JSON.
• /agents/ai_decisions_gateway/skill.md — Markdown for coding agents.