Building an MCP Server for Java Performance Profiling

#java #performance #mcp

I built javaperf - an MCP (Model Context Protocol) server that lets AI assistants profile Java applications without touching the command line.

If you've ever debugged Java performance issues, you know the drill: jps, jcmd, jfr... memorizing flags, parsing output. What if your AI assistant could do this for you?

What is MCP?

Model Context Protocol (MCP) is a standard for connecting AI assistants to external tools. Instead of AI just generating code suggestions, it can actually:

Run diagnostics on your running Java apps
Analyze thread dumps
Inspect JFR recordings
All through natural conversation

javaperf Features

Process Discovery: Auto-detect running Java processes via jps
Thread Analysis: Dump and analyze threads (jcmd Thread.print)
VM Flags: Inspect JVM configuration
JFR Support: Record and analyze Java Flight Recorder data
Performance Counters: Access low-level JVM metrics

Installation

npm install -g javaperf

Or use directly via npx (no install needed):

npx javaperf

Connecting to Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "javaperf": {
      "command": "npx",
      "args": ["-y", "javaperf"]
    }
  }
}

Example Usage

Once connected, you can ask Claude things like:

"Show me all running Java processes"
"Analyze thread dump for process 1234"
"Start a 60-second JFR recording"
"Why is my app using so much memory?"

The AI handles the jcmd, jfr, and jps commands behind the scenes.

Why This Matters

Before MCP: You Google the right flags, copy-paste commands, manually parse output.

After MCP: "Hey Claude, why is my Spring Boot app slow?" → AI investigates → Finds thread contention → Suggests fixes.

Try It Out

If you're working with Java and using Claude Desktop or Cursor IDE, give it a spin. Would love feedback on what other profiling features would be useful!

Built with TypeScript, powered by JDK tools you already have.

DEV Community