简体中文 | English

Pprof Analyzer MCP Server

This is a Model Context Protocol (MCP) server implemented in Go, providing a tool to analyze Go pprof performance profiles.

Features

• analyze_pprof Tool:
  • Analyzes the specified Go pprof file and returns serialized analysis results (e.g., Top N list or flame graph JSON).
  • Supported Profile Types:
    • cpu: Analyzes CPU time consumption during code execution to find hot spots.
    • heap: Analyzes the current memory usage (heap allocations) to find objects and functions with high memory consumption. Enhanced with object count, allocation site, and type information.
    • goroutine: Displays stack traces of all current goroutines, used for diagnosing deadlocks, leaks, or excessive goroutine usage.
    • allocs: Analyzes memory allocations (including freed ones) during program execution to locate code with frequent allocations. Provides detailed allocation site and object count information.
    • mutex: Analyzes contention on mutexes to find locks causing blocking. (Not yet implemented)
    • block: Analyzes operations causing goroutine blocking (e.g., channel waits, system calls). (Not yet implemented)
  • Supported Output Formats: text, markdown, json (Top N list), flamegraph-json (hierarchical flame graph data, default).
    • text, markdown: Human-readable text or Markdown format.
    • json: Outputs Top N results in structured JSON format (implemented for cpu, heap, goroutine, allocs).
    • flamegraph-json: Outputs hierarchical flame graph data in JSON format, compatible with d3-flame-graph (implemented for cpu, heap, allocs, default format). Output is compact.
  • Configurable number of Top N results (top_n, defaults to 5, effective for text, markdown, json formats).
• generate_flamegraph Tool:
  • Uses go tool pprof to generate a flame graph (SVG format) for the specified pprof file, saves it to the specified path, and returns the path and SVG content.
  • Supported Profile Types: cpu, heap, allocs, goroutine, mutex, block.
  • Requires the user to specify the output SVG file path.
  • Important: This feature depends on Graphviz being installed.
• open_interactive_pprof Tool (macOS Only):
  • Attempts to launch the go tool pprof interactive web UI in the background for the specified pprof file. Uses port :8081 by default if http_address is not provided.
  • Returns the Process ID (PID) of the background pprof process upon successful launch.
  • macOS Only: This tool will only work on macOS.
  • Dependencies: Requires the go command to be available in the system's PATH.
  • Limitations: Errors from the background pprof process are not captured by the server. Temporary files downloaded from remote URLs are not automatically cleaned up until the process is terminated (either manually via disconnect_pprof_session or when the MCP server exits).
• detect_memory_leaks Tool:
  • Compares two heap profile snapshots to identify potential memory leaks.
  • Analyzes memory growth by object type and allocation site.
  • Provides detailed statistics on memory growth, including absolute and percentage changes.
  • Configurable growth threshold and result limit.
  • Helps identify memory leaks by comparing profiles taken at different points in time.
• disconnect_pprof_session Tool:
  • Attempts to terminate a background pprof process previously started by open_interactive_pprof, using its PID.
  • Sends an Interrupt signal first, then a Kill signal if Interrupt fails.

Installation (As a Library/Tool)

You can install this package directly using go install:

go install github.com/ZephyrDeng/pprof-analyzer-mcp@latest

This will install the pprof-analyzer-mcp executable to your $GOPATH/bin or $HOME/go/bin directory. Ensure this directory is in your system's PATH to run the command directly.

Building from Source

Ensure you have a Go environment installed (Go 1.18 or higher recommended).

In the project root directory (pprof-analyzer-mcp), run:

go build

This will generate an executable file named pprof-analyzer-mcp (or pprof-analyzer-mcp.exe on Windows) in the current directory.

Using go install (Recommended)

You can also use go install to install the executable into your $GOPATH/bin or $HOME/go/bin directory. This allows you to run pprof-analyzer-mcp directly from the command line (if the directory is added to your system's PATH environment variable).

# Installs the executable using the module path defined in go.mod
go install .
# Or directly using the GitHub path (recommended after publishing)
# go install github.com/ZephyrDeng/pprof-analyzer-mcp@latest

Running with Docker

Using Docker is a convenient way to run the server, as it bundles the necessary Graphviz dependency.

1. Build the Docker Image: In the project root directory (where the Dockerfile is located), run:

   docker build -t pprof-analyzer-mcp .
   

2. Run the Docker Container:

   docker run -i --rm pprof-analyzer-mcp
   

   • The -i flag keeps STDIN open, which is required for the stdio transport used by this MCP server.
   • The --rm flag automatically removes the container when it exits.

3. Configure MCP Client for Docker: To connect your MCP client (like Roo Cline) to the server running inside Docker, update your .roo/mcp.json:

   {
     "mcpServers": {
       "pprof-analyzer-docker": {
         "command": "docker run -i --rm pprof-analyzer-mcp"
       }
     }
   }
   

   Make sure the pprof-analyzer-mcp image has been built locally before the client tries to run this command.

Releasing (Automated via GitHub Actions)

This project uses GoReleaser and GitHub Actions to automate the release process. Releases are triggered automatically when a Git tag matching the pattern v* (e.g., v0.1.0, v1.2.3) is pushed to the repository.

Release Steps:

1. Make Changes: Develop new features or fix bugs.
2. Commit Changes: Commit your changes using Conventional Commits format (e.g., feat: ..., fix: ...). This is important for automatic changelog generation.

   git add .
   git commit -m "feat: Add awesome new feature"
   # or
   git commit -m "fix: Resolve issue #42"
   

3. Push Changes: Push your commits to the main branch on GitHub.

   git push origin main
   

4. Create and Push Tag: When ready to release, create a new Git tag and push it to GitHub.

   # Example: Create tag v0.1.0
   git tag v0.1.0
   
   # Push the tag to GitHub
   git push origin v0.1.0
   

5. Automatic Release: Pushing the tag will trigger the GoReleaser GitHub Action defined in .github/workflows/release.yml. This action will:
   • Build binaries for Linux, macOS, and Windows (amd64 & arm64).
   • Generate a changelog based on Conventional Commits since the last tag.
   • Create a new GitHub Release with the changelog and attach the built binaries and checksums as assets.

You can view the release workflow progress in the "Actions" tab of the GitHub repository.

Configuring the MCP Client

This server uses the stdio transport protocol. You need to configure it in your MCP client (e.g., Roo Cline extension for VS Code).

Typically, this involves adding the following configuration to the .roo/mcp.json file in your project root:

{
  "mcpServers": {
    "pprof-analyzer": {
      "command": "pprof-analyzer-mcp"
    }
  }
}

Note: Adjust the command value based on your build method (go build or go install) and the actual location of the executable. Ensure the MCP client can find and execute this command.

After configuration, reload or restart your MCP client, and it should automatically connect to the PprofAnalyzer server.

Dependencies

• Graphviz: The generate_flamegraph tool requires Graphviz to generate SVG flame graphs (the go tool pprof command calls dot when generating SVG). Ensure Graphviz is installed on your system and the dot command is available in your system's PATH environment variable.

  Installing Graphviz:

  • macOS (using Homebrew):

    brew install graphviz
    

  • Debian/Ubuntu:

    sudo apt-get update && sudo apt-get install graphviz
    

  • CentOS/Fedora:

    sudo yum install graphviz
    # or
    sudo dnf install graphviz
    

  • Windows (using Chocolatey):

    choco install graphviz
    

  • Other Systems: Refer to the Graphviz official download page.

Usage Examples (via MCP Client)

Once the server is connected, you can call the analyze_pprof and generate_flamegraph tools using file://, http://, or https:// URIs for the profile file.

Example: Analyze CPU Profile (Text format, Top 5)

{
  "tool_name": "analyze_pprof",
  "arguments": {
    "profile_uri": "file:///path/to/your/cpu.pprof",
    "profile_type": "cpu"
  }
}

Example: Analyze Heap Profile (Markdown format, Top 10)

{
  "tool_name": "analyze_pprof",
  "arguments": {
    "profile_uri": "file:///path/to/your/heap.pprof",
    "profile_type": "heap",
    "top_n": 10,
    "output_format": "markdown"
  }
}

Example: Analyze Goroutine Profile (Text format, Top 5)

{
  "tool_name": "analyze_pprof",
  "arguments": {
    "profile_uri": "file:///path/to/your/goroutine.pprof",
    "profile_type": "goroutine"
  }
}

Example: Generate Flame Graph for CPU Profile

{
  "tool_name": "generate_flamegraph",
  "arguments": {
    "profile_uri": "file:///path/to/your/cpu.pprof",
    "profile_type": "cpu",
    "output_svg_path": "/path/to/save/cpu_flamegraph.svg"
  }
}

Example: Generate Flame Graph for Heap Profile (inuse_space)

{
  "tool_name": "generate_flamegraph",
  "arguments": {
    "profile_uri": "file:///path/to/your/heap.pprof",
    "profile_type": "heap",
    "output_svg_path": "/path/to/save/heap_flamegraph.svg"
  }
}

Example: Analyze CPU Profile (JSON format, Top 3)

{
  "tool_name": "analyze_pprof",
  "arguments": {
    "profile_uri": "file:///path/to/your/cpu.pprof",
    "profile_type": "cpu",
    "top_n": 3,
    "output_format": "json"
  }
}

Example: Analyze CPU Profile (Default Flame Graph JSON format)

{
  "tool_name": "analyze_pprof",
  "arguments": {
    "profile_uri": "file:///path/to/your/cpu.pprof",
    "profile_type": "cpu"
    // output_format defaults to "flamegraph-json"
  }
}

Example: Analyze Heap Profile (Explicitly Flame Graph JSON format)

{
  "tool_name": "analyze_pprof",
  "arguments": {
    "profile_uri": "file:///path/to/your/heap.pprof",
    "profile_type": "heap",
    "output_format": "flamegraph-json"
  }
}

Example: Analyze Remote CPU Profile (from HTTP URL)

{
  "tool_name": "analyze_pprof",
  "arguments": {
    "profile_uri": "https://example.com/profiles/cpu.pprof",
    "profile_type": "cpu"
  }
}

Example: Analyze Online CPU Profile (from GitHub Raw URL)

{
  "tool_name": "analyze_pprof",
  "arguments": {
    "profile_uri": "https://raw.githubusercontent.com/google/pprof/refs/heads/main/profile/testdata/gobench.cpu",
    "profile_type": "cpu",
    "top_n": 5
  }
}

Example: Generate Flame Graph for Online Heap Profile (from GitHub Raw URL)

{
  "tool_name": "generate_flamegraph",
  "arguments": {
    "profile_uri": "https://raw.githubusercontent.com/google/pprof/refs/heads/main/profile/testdata/gobench.heap",
    "profile_type": "heap",
    "output_svg_path": "./online_heap_flamegraph.svg"
  }
}

Example: Open Interactive Pprof UI for Online CPU Profile (macOS Only)

{
  "tool_name": "open_interactive_pprof",
  "arguments": {
    "profile_uri": "https://raw.githubusercontent.com/google/pprof/refs/heads/main/profile/testdata/gobench.cpu"
    // Optional: "http_address": ":8082" // Example of overriding the default port
  }
}

Example: Detect Memory Leaks Between Two Heap Profiles

{
  "tool_name": "detect_memory_leaks",
  "arguments": {
    "old_profile_uri": "file:///path/to/your/heap_before.pprof",
    "new_profile_uri": "file:///path/to/your/heap_after.pprof",
    "threshold": 0.05,  // 5% growth threshold
    "limit": 15         // Show top 15 potential leaks
  }
}

Example: Disconnect a Pprof Session

{
  "tool_name": "disconnect_pprof_session",
  "arguments": {
    "pid": 12345 // Replace 12345 with the actual PID returned by open_interactive_pprof
  }
}

Future Improvements (TODO)

• Implement full analysis logic for mutex, block profiles.
• Implement json output format for mutex, block profile types.
• Set appropriate MIME types in MCP results based on output_format.
• Add more robust error handling and logging level control.
• Consider supporting remote pprof file URIs (e.g., http://, https://). (Done)
• Implement full analysis logic for allocs profiles. (Done)
• Implement json output format for allocs profile type. (Done)
• Add memory leak detection capabilities. (Done)
• Add time-series analysis for memory profiles to track growth over multiple snapshots.
• Implement differential flame graphs to visualize changes between profiles.