Result mapping

Anvia maps MCP responses to the string result used by ordinary tools.

1. Map returned content

Text content is returned directly. Multiple content items are concatenated in server order.

Image content becomes a structured file content part: { type: 'file', data: { type: 'data', data }, mediaType }. The base64 payload stays data; it is not wrapped in a data: URL.

Text and binary resources retain their media type, URI, and text or blob in a serialized string.

audio and resource_link content are unsupported: the adapted tool throws Unsupported MCP tool result content type.

When the server returns empty content alongside structuredContent, the structured value is JSON-serialized into a single text part.

A response containing { toolResult } returns the string directly or JSON-serializes a non-string value when possible.

{
  content: [
    { type: 'text', text: 'The deployment is healthy.' },
  ],
}

The adapted tool returns The deployment is healthy.

2. Handle MCP error results

When isError: true, the adapted tool throws. Text content becomes the error message; without text, the message is MCP tool returned an error.

A direct tool.call() therefore rejects. Inside an agent run, normal tool execution catches the failure, records a failed tool event, and returns the error text to the model as a tool result unless another runtime boundary fails the run.

Do not expose raw MCP error text to a browser. It may include server names, paths, arguments, or upstream details.

3. Validate arguments

MCP tool calls accept a JSON object. null or undefined omits the arguments field. Arrays, primitives, and other non-object values reject before the remote call.

The remote server must still validate arguments and authorize the action. A model-facing input schema guides generation; it is not server-side enforcement.

4. Bound untrusted output

Remote text, images, and resources may be private, stale, malicious, or oversized. Use middleware or an app-owned wrapper to truncate, redact, store large artifacts, or create a smaller model-facing result.

Next, enforce MCP trust boundaries.