Skip to main content
OpenAI tool protection has two separate stages:
  • Tool exposure: remove tools before the model sees them.
  • Tool-call check: evaluate a model-requested tool call before your app executes it.
Do not collapse these in your mental model. Exposure controls planning. Tool-call checks control execution.

​
Responses Tools

For Responses API request-time filtering, the wrapper understands:
  • function tools, keyed by name
  • custom tools, keyed by name
  • hosted OpenAI tools, keyed by type
  • MCP tools, keyed as mcp:<server_label> when server_label exists
const tools = [
  {
    type: "function" as const,
    name: "search_docs",
    description: "Search docs.",
    strict: true,
    parameters: {
      type: "object",
      properties: {
        query: { type: "string" },
      },
      required: ["query"],
      additionalProperties: false,
    },
  },
  {
    type: "function" as const,
    name: "send_email",
    description: "Send an email.",
    strict: true,
    parameters: {
      type: "object",
      properties: {
        to: { type: "string" },
        subject: { type: "string" },
        body: { type: "string" },
      },
      required: ["to", "subject", "body"],
      additionalProperties: false,
    },
  },
];
If policy blocks send_email, the wrapper forwards only the remaining tools.

​
Chat Completions Tools

For Chat Completions, the wrapper understands function and custom tools. 
const tools = [
  {
    type: "function" as const,
    function: {
      name: "search_docs",
      description: "Search docs.",
      strict: true,
      parameters: {
        type: "object",
        properties: {
          query: { type: "string" },
        },
        required: ["query"],
        additionalProperties: false,
      },
    },
  },
];
When tools are filtered, the wrapper also sanitizes tool_choice so the request does not force a blocked tool.

​
Decision Callback

Use onDecision to inspect original, blocked, and forwarded tools. 
client = wrapOpenAI(client, {
  onDecision(event) {
    if (event.checkpointType !== "request") {
      return;
    }

    console.log("decision:", event.decision.decision);
    console.log("blocked tools:", event.decision.blockedTools ?? []);
    console.log("forwarded tools:", event.forwardedTools.length);
  },
});

​
Tool-call Checks

After OpenAI returns a tool call, Averta checks it before guarded tool-call events are exposed to your app. If the tool-call decision is block, the wrapper throws. Your app should not execute the tool.

​
Debugging

SymptomCheck
Tool disappearedInspect request decision blockedTools.
Forced tool choice was ignoredThe requested tool may have been blocked and tool_choice sanitized.
Tool-call event throwsThe tool-call checkpoint blocked execution.
Hosted tool not filtered as expectedConfirm policy uses the normalized tool identity, such as web_search_preview or mcp:<server_label>.