How AI-Powered Arthas with MCP Transforms Online Issue Diagnosis

Introduction

Developers often need to diagnose online Java issues such as CPU spikes, slow interfaces, or memory leaks without restarting the application.

Arthas – Java "Swiss‑army knife"

Arthas uses the Java Instrumentation API and ASM bytecode enhancement to inspect threads, trace methods, monitor parameters, and view class‑loading information at runtime. Executing trace com.example.Service method inserts temporary bytecode at method entry and exit to measure execution time; the modifications are automatically reverted when Arthas exits.

Model Context Protocol (MCP)

MCP, open‑sourced by Anthropic in November 2024, standardises the JSON‑RPC 2.0 interface for AI assistants to invoke tools. As of early 2026, more than 10,000 MCP servers are active in production with monthly SDK downloads of 97 million.

The MCP architecture consists of an MCP Host (the AI‑enabled application, e.g., Claude Desktop or Cursor), an MCP Client that communicates with the host, and an MCP Server that provides diagnostic capabilities. Communication supports both stdio and HTTP transport modes.

Arthas + MCP Integration

The experimental Arthas MCP Server module exposes Arthas diagnostic commands through an HTTP/Netty JSON‑RPC 2.0 endpoint, allowing AI to invoke them directly. Two protocol modes are available:

STREAMABLE : Server‑Sent Events (SSE) push streaming responses for real‑time monitoring.

STATELESS : Traditional request‑response suitable for simple integration.

Documentation: https://arthas.aliyun.com/doc/mcp-server.html

Practical Case Study

Scenario: An order‑service experiences 98 % CPU usage during peak hours, causing slow responses and user complaints.

"My order service CPU spiked to 98 %, help me identify the thread causing high CPU and where it is stuck."

AI follows a built‑in troubleshooting script:

Invoke dashboard to list threads. Sample output:

ID   NAME                     CPU%   STATE
29   http-nio-8080-exec-8     89.2   RUNNABLE
12   DubboServerHandler-...   3.1    RUNNABLE
...

Thread 29 (http‑nio‑8080‑exec‑8) is identified as the culprit.

Invoke thread with the thread ID to retrieve the full stack trace. Sample output includes a call chain ending at com.example.order.service.OrderService.getOrder and a regex match in java.util.regex.Pattern.

Analyze the stack: the thread is blocked in a greedy .* regular expression inside a logging aspect, causing catastrophic backtracking when processing large JSON payloads.

Problem location: LogAspect.java:47 regex. Root cause: Greedy .* pattern on >10 KB JSON triggers backtracking.

AI‑generated remediation suggestions:

Temporarily disable the logging aspect via configuration.

Replace .* with a non‑greedy pattern such as [^"]* or use String.indexOf instead of regex.

Add a length limit to the aspect (e.g., skip sanitisation for strings >2000 characters).

Pros and Cons

No need to learn Arthas command syntax; natural language suffices.

Unified integration with AI clients (Claude Desktop, Cursor, Cherry Studio) via MCP.

Secure access through Bear Token authentication.

Includes 26 core diagnostic tools covering JVM monitoring, class loading, and method tracing.

HTTP‑based, easy to embed in IDEs or monitoring systems.

Experimental module; features are rapidly evolving.

Requires manual MCP client configuration.

Complex scenarios may still need human verification of AI conclusions.

Applicable Scenarios

Rapid daily online fault diagnosis.

Assisting junior engineers during onboarding.

Code‑level issue localisation within IDEs.

Building internal AI‑assisted operations assistants.

Conclusion

Arthas’s integration with MCP enables AI‑assisted online troubleshooting: developers describe symptoms in plain language, the AI automatically executes diagnostics, analyses results, and proposes fixes, reducing the cognitive load of remembering commands and accelerating incident resolution, especially for newcomers.