Skybridge

Skybridge builds MCP Apps for ChatGPT, Claude and other MCP clients. autotel-skybridge adds one middleware that traces every tool call, resource read and prompt, with the spans and attributes of autotel-mcp-instrumentation.

Installation

npm install autotel autotel-skybridge

Usage

Init autotel in its own module and import it first in src/index.ts:

src/instrumentation.ts

import { init } from 'autotel';

init({ service: 'my-app' }); // reads OTEL_EXPORTER_OTLP_ENDPOINT / _HEADERS

Register the middleware on the server:

import { skybridgeTracing } from 'autotel-skybridge';
import { Skybridge } from 'skybridge/server';

export const app = new Skybridge({
  name: 'my-app',
  version: '1.0.0',
  handler: (server) =>
    server
      .registerTool({ name: 'search' /* ... */ }, handler)
      .mcpMiddleware(skybridgeTracing({ captureToolArgs: true })),
});

Spans

Request	Span name	Key attributes
tools/call	tools/call <tool>	gen_ai.tool.name, mcp.method.name
resources/read	resources/read	mcp.resource.uri (no ?v= cache key)
prompts/get	prompts/get <prompt>	gen_ai.prompt.name

A call to an unknown tool gets the bare method name, with no tool label on the duration metric. Resource URIs stay on the span and off the metric, since a URI can expand a template.

The caller’s traceparent in _meta parents each span, so an instrumentMcpClient caller and your server share one trace. Spans you start inside a tool nest under the tool span.

When a tool throws, Skybridge answers with an isError result. The span records ERROR with the failure text as its status message, error.type=tool_error, mcp.failure.category and mcp.failure.fingerprint, and the original exception. A createStructuredError keeps its why, fix and code as error.* attributes.

Deploying to Vercel

skybridge build writes Vercel’s prebuilt output. Flush per request and keep the bundled JSON exporter:

init({
  service: 'my-app',
  protocol: 'http',
  forceFlushOnShutdown: process.env.VERCEL === '1',
});

Set OTEL_EXPORTER_OTLP_ENDPOINT and OTEL_EXPORTER_OTLP_HEADERS on the Vercel project and run vercel deploy --prebuilt.

Skybridge sends its own usage telemetry unless SKYBRIDGE_TELEMETRY_DISABLED=1 is set.

Example

apps/example-skybridge has a tool with a React view, a tool that throws, and a traced client script.