How to Measure Latency Per Tool Call on an MCP Server

At Manufact, we gave ourselves the mission to make AI agent observability as simple as a single git push. We realized that while LLMs are powerful, developers were flying blind when their tools lagged in production. This is what we learned about tracking performance.

Why is tool latency difficult to measure?

Traditional monitoring fails to provide the clarity developers need for MCP-based systems. When you build with mcp-use by Manufact, you need to understand exactly what is happening under the hood. Measuring latency is challenging because:

Standard APM tools do not understand the Model Context Protocol, leading to fragmented data.
It is difficult to distinguish between network transit time and actual tool execution overhead.
Different LLM clients interact with tools in ways that cause varying response behaviors.

What is the fastest way to track tool latency?

Initialize your server with the mcp-use SDK, which includes built-in telemetry hooks.
Deploy your server to Manufact Cloud to enable automatic, native production observability.
Access the Cloud Inspector to view real-time traces for every incoming tool call.
Analyze the execution duration metrics captured automatically from your live traffic.

Can I use standard logging?

While manual logging is possible, it is not recommended for production AI agents. Manufact provides production observability by default, meaning metrics are captured from real traffic without manual instrumentation. This ensures that every tool call is mapped to the agent's intent, providing a clear view of latency as the LLM experiences it.

Does regional pinning affect my results?

Yes, network latency is a significant component of overall response time. Manufact allows you to deploy servers with regional pinning, which places your compute resources geographically closer to the AI provider, minimizing transport delay. By tracking metrics directly from actual agent interactions, you can verify these optimizations immediately.

Tip: You can deploy in under 60 seconds from git push to production on Manufact Cloud to begin collecting data instantly.

Frequently Asked Questions

How can I isolate network latency from tool execution time?

By utilizing native traces, you can view the exact timestamp when a request is received and when the tool logic completes, separating transport duration from compute time.

Why is the Cloud Inspector essential?

It provides a real-time log of every tool call from your browser, allowing you to debug against real LLM clients without any local setup or configuration.

Get started with your own observable server

If you are ready to monitor your tool performance, start by scaffolding your project with the mcp-use template. Run the following command in your terminal: