Transform Responses

Middleware can inspect the upstream response and return a new value. Check the compatibility result shape before spreading content.

Add source context

import { hasToolContent } from 'mcpose';
import type { ToolMiddleware } from 'mcpose';

export const withSource: ToolMiddleware = async (req, next) => {
  const result = await next(req);
  if (!hasToolContent(result)) return result;
  return {
    ...result,
    content: [...result.content, { type: 'text', text: 'Source: internal docs' }],
  };
};

Configure a named proxy with toolMiddleware: [withSource]. The source annotation is a transformation, not a security filter.

Make every content decision explicit

For sanitization, text is only one channel: tool results can also contain images, embedded resources, and structured JSON. mapToolResult requires a handler for every channel. This example redacts text and deliberately drops non-text blocks and structured content.

import { mapToolResult } from 'mcpose';
import type { ToolMiddleware } from 'mcpose';

export const redact: ToolMiddleware = async (req, next) =>
  mapToolResult(await next(req), {
    onText: block => ({ ...block, text: block.text.replace(/demo@example\.test/g, '[REDACTED]') }),
    onOther: () => null,
    onStructured: () => undefined,
  });

Legacy { toolResult } compatibility responses and unknown extra fields pass through unchanged. Decide separately whether to reject or sanitize those shapes at your boundary. The demonstration pattern is not a general PII detector. Define your own data classification and transformation rules, and verify their behavior with representative payloads. To audit transformed responses, place audit outside the transformer: [redact, audit.middleware]. Audit captures request arguments on entry, so response ordering alone does not sanitize input arguments.