Tool use

Anthropic tool protection has two separate stages:

Tool exposure: remove blocked tools before Anthropic sees the request.
Tool-call check: evaluate returned tool_use blocks before your app executes local code.

Keep those separate. Exposure controls what the model can plan with. Tool-call checks control what your app is allowed to execute.

Supported Tools

The current wrapper supports user-defined client tools with a stable name. The tool type may be omitted, custom, or function.

const tools = [
  {
    name: "search_docs",
    description: "Search internal support documentation.",
    input_schema: {
      type: "object",
      properties: {
        query: { type: "string" },
      },
      required: ["query"],
      additionalProperties: false,
    },
  },
  {
    name: "send_email",
    description: "Send an email to a customer.",
    input_schema: {
      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.

Tool Choice

When tools are filtered, the wrapper also sanitizes tool_choice. If the request forces a specific tool with tool_choice: { type: "tool", name: "send_email" } and Averta blocks send_email, the wrapper removes that forced choice before forwarding the request. Otherwise the request would ask Anthropic to call a tool it can no longer see.

Tool-call Checks

After Anthropic responds, the wrapper reads each returned tool_use block:

const toolUses = response.content.filter((block) => block.type === "tool_use");

For each block, Averta checks:

the tool call id from tool_use.id
the tool name from tool_use.name
the tool arguments from tool_use.input
bounded conversation context from the current Messages request
the request run_id created by the request checkpoint

If a tool-call decision is block, the wrapper throws AvertaSdkError and does not return that response to your application. Your app should not execute the tool.

Decision Callback

client = wrapAnthropic(client, {
  onDecision(event) {
    if (event.checkpointType === "request") {
      console.log("blocked tools:", event.decision.blockedTools ?? []);
      console.log("forwarded tools:", event.forwardedTools.length);
    }

    if (event.checkpointType === "tool_call") {
      console.log(event.tool.name, event.decision.decision);
    }
  },
});

Unsupported Tool Shapes

Unsupported Anthropic tool shapes fail closed before provider execution. This is deliberate. A guardrail wrapper that silently forwards unsupported tools would create false confidence.

Debugging

Symptom	Check
A tool disappeared	Inspect request decision `blockedTools`.
Forced tool choice was ignored	The forced tool may have been blocked and `tool_choice` sanitized.
`messages.create(...)` throws before Anthropic is called	Check for unsupported tool shapes or streaming input.
Tool execution never starts	The returned `tool_use` may have been blocked by the tool-call checkpoint.

Get Started

Core Concepts

OpenAI

Anthropic

Dashboard

Advanced

Supported Tools

Tool Choice

Tool-call Checks

Decision Callback

Unsupported Tool Shapes

Debugging

​Supported Tools

​Tool Choice

​Tool-call Checks

​Decision Callback

​Unsupported Tool Shapes

​Debugging

Supported Tools

Tool Choice

Tool-call Checks

Decision Callback

Unsupported Tool Shapes

Debugging