Perfetto MCP

Turn natural language into powerful Perfetto trace analysis

A Model Context Protocol (MCP) server that transforms natural-language prompts into focused Perfetto analyses. Quickly explain jank, diagnose ANRs, spot CPU hot threads, uncover lock contention, and find memory leaks – all without writing SQL.

✨ Features

• Natural Language → SQL: Ask questions in plain English, get precise Perfetto queries
• ANR Detection: Automatically identify and analyze Application Not Responding events
• Performance Analysis: CPU profiling, frame jank detection, memory leak detection
• Thread Contention: Find synchronization bottlenecks and lock contention
• Binder Profiling: Analyze IPC performance and slow system interactions

🛠️ Available Tools

🔎 Exploration & Discovery

🚨 ANR Analysis

Note: Helpful if the recorded trace contains ANR

🎯 Performance Profiling

📱 UI Performance

🔒 Concurrency & IPC

💾 Memory Analysis

Output Format

All tools return structured JSON with:

• Summary: High-level findings
• Details: Tool-specific results
• Metadata: Execution context and any fallbacks used

📚 Resources

• Trace Processor Python API - Perfetto's Python interface
• Perfetto SQL Syntax - SQL reference for custom queries

📄 License

Apache 2.0 License. See LICENSE for details.

📋 Prerequisites

• Python 3.13+ (macOS/Homebrew):
  Copy & paste — that's itCopy

  brew install python@3.13

• uv (recommended):
  Copy & paste — that's itCopy

  brew install uv

🚀 Getting Started

Or add to ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "perfetto-mcp": {
      "command": "uvx",
      "args": ["perfetto-mcp"]
    }
  }
}

Run this command. See Claude Code MCP docs for more info.

# Add to user scope
claude mcp add perfetto-mcp --scope user -- uvx perfetto-mcp

Or edit ~/claude.json (macOS) or %APPDATA%\Claude\claude.json (Windows):

{
  "mcpServers": {
    "perfetto-mcp": {
      "command": "uvx",
      "args": ["perfetto-mcp"]
    }
  }
}

<img alt="Install in VS Code" src="https://img.shields.io/badge/VS_Code-VS_Code?style=flat-square&label=Install%20Perfetto%20MCP&color=0098FF">

or add to .vscode/mcp.json (project) or run "MCP: Add Server" command:

{
  "mcpServers": {
    "perfetto-mcp": {
      "command": "uvx",
      "args": ["perfetto-mcp"]
    }
  }
}

Enable in GitHub Copilot Chat's Agent mode.

Edit ~/.codex/config.toml:

[mcp_servers.perfetto-mcp]
command = "uvx"
args = ["perfetto-mcp"]

Optional: Use a Local trace_processor_shell Binary

If your network environment blocks downloads, set PERFETTO_MCP_TRACE_PROCESSOR_BIN_PATH to an absolute path of a local trace_processor_shell binary.

When this env var is set, perfetto-mcp uses that binary directly. When it is not set, default perfetto Python behavior is unchanged.

Example (mcp.json):

{
  "mcpServers": {
    "perfetto-mcp": {
      "command": "uvx",
      "args": ["perfetto-mcp"],
      "env": {
        "PERFETTO_MCP_TRACE_PROCESSOR_BIN_PATH": "D:/tools/perfetto/trace_processor_shell.exe"
      }
    }
  }
}

Example (~/.codex/config.toml):

[mcp_servers.perfetto-mcp]
command = "uvx"
args = ["perfetto-mcp"]
[mcp_servers.perfetto-mcp.env]
PERFETTO_MCP_TRACE_PROCESSOR_BIN_PATH = "D:/tools/perfetto/trace_processor_shell.exe"

Local Install (development server)

cd perfetto-mcp-server
uv sync
uv run mcp dev src/perfetto_mcp/dev.py

{
  "mcpServers": {
    "perfetto-mcp-local": {
      "command": "uv",
      "args": [
        "--directory",
        "/path/to/git/repo/perfetto-mcp",
        "run",
        "-m",
        "perfetto_mcp"
      ],
      "env": { "PYTHONPATH": "src" }
    }
  }
}

pip3 install perfetto-mcp
python3 -m perfetto_mcp

📖 How to Use

Example starting prompt:

In the perfetto trace, I see that the FragmentManager is taking 438ms to execute. Can you figure out why it's taking so long?

Required Parameters

Every tool needs these two inputs:

In Your Prompts

Be explicit about the trace and process, prefix your prompt with:

"Use perfetto trace /absolute/path/to/trace.perfetto-trace for process com.example.app"

Optional Filters

Many tools support additional filtering (but let your LLM handle that):

• time_range: {start_ms: 10000, end_ms: 25000}
• Tool-specific thresholds: min_block_ms, jank_threshold_ms, limit