Skip to main content
Glama
profiler-mcp

profiler-mcp

by Lagooneng

profiler-mcp

English | 한국어

An MCP server that measures where a running C++ process spends its time: the functions using CPU, time spent waiting, and GPU usage, aggregated down to the function level with PDBs. The target program does not need to be modified or restarted. Tools for Unreal Engine are included.

Requirements

Item

Install

Windows 10 or later (64-bit)

—

uv

winget install astral-sh.uv

msdia140.dll

Visual Studio's "Desktop development with C++" workload

Very Sleepy (optional)

winget install codersnotes.VerySleepy — only needed for quick_sample

Unreal Engine (optional)

The ue_ tools are enabled when it is installed

Related MCP server: wpa-mcp

Install (Claude Code plugin)

/plugin marketplace add Lagooneng/profiler-mcp
/plugin install profiler-mcp@profiler-marketplace

What to measure with

What you want to know

Use

What is heavy right now, quickly

quick_sample → top_functions / call_tree

How many ms, and how much a fix saved

capture_start / capture_stop before and after → compare

Which thread is busy

capture → threads

What it is waiting on

capture_start(mode='cpu+wait') → wait_summary

Whether the GPU is the bottleneck

capture → gpu_usage

When frames hitch

capture_start(gpu=True) → gpu_frames

Unreal: engine scopes, frames, render passes

ue_frames, ue_top_timers, ue_timer_callees

  • quick_sample samples one thread for a few seconds without administrator rights. The target slows down noticeably while it runs, so use only the ratios and ranking. It works by running Very Sleepy (GPL-2.0) and reading its result file. Because of that license, this repository does not include its executable or code; it uses the copy you install yourself.

  • capture_start / capture_stop record the whole system with ETW. A UAC prompt appears once for start and once for stop.

  • For Unreal projects, Unreal Insights traces (the ue_ tools) are the default. They need no administrator rights.

Tools

Tool

What it does

status

Tool paths, capture in progress, storage

capture_start / capture_stop / capture_cancel

Start an ETW capture · stop and save · discard

quick_sample

Sample one thread without administrator rights

list_traces / delete_trace

Manage saved traces

processes / threads

CPU usage per process · per thread

trace_info

Capture quality (stack coverage, modules that had PDBs)

top_functions / callers_callees / call_tree

Time per function, call relationships

compare

Per-function change between two captures

wait_summary

Waiting time by reason · by the thread that woke it

gpu_frames / gpu_usage

Frame presentation intervals, GPU usage per process

ue_trace_commands / ue_list_traces / ue_delete_trace

Record and manage Insights traces

ue_frames / ue_threads

Frame times, thread list

ue_top_timers / ue_timer_callees

Time per engine scope (CPU threads, GPU queues)

ue_sampling_window

Where an ETW capture and an Insights trace overlap

The profile-report skill is installed with the plugin: the format for reporting results and the checks to run before reporting.

Good to know

  • The UAC prompt shows "Windows Performance Recorder". Any other name is not a prompt from this server.

  • An ETW capture records the whole system. The trace contains the names and stacks of other processes. Take care when sharing the file.

  • A capture keeps running independently of this server. Once started, stop it or cancel it.

  • Modules without a PDB are grouped as [module name]. Symbols for the kernel and Windows system DLLs are not supported.

  • Storage: a single capture overwrites its oldest part at 10 GB, and when saved traces exceed 20 GB in total the least recently used are deleted.

Environment variables

Variable

Meaning

PROFILER_CACHE_DIR

Where traces and PDB copies are kept (default %LOCALAPPDATA%\profiler-mcp)

PROFILER_MAX_CAPTURE_GB

Size cap for a single capture (default 10, 0 for no cap)

PROFILER_TRACES_CAP_GB

Cap on the total size of saved traces (default 20, 0 to never clean up)

PROFILER_SYMBOL_DIRS

Folders of PDBs that are never rebuilt (separated by ;)

PROFILER_WPR_PATH / PROFILER_MSDIA_PATH / PROFILER_SLEEPY_PATH / PROFILER_UNREAL_INSIGHTS_PATH

Paths to tools that are not found in their default location

PROFILER_SKIP_BOOT_CHECK

1 skips the check that an ETW capture and an Insights trace come from the same boot

Tests

uv run python tests\test_units.py
uv run python tests\test_mcp_e2e.py <process name> [trace name] [--capture]
uv run python tests\test_unreal_insights.py [trace name]

--capture runs a real capture, so the UAC prompt appears twice.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      Not graded
      maintenance
      Enables GPU profiling and performance analysis via NVIDIA Nsight Systems, allowing agents to profile binaries and aggregate statistics for kernels, memory copies, and NVTX ranges. It supports advanced analysis through interval tree construction and structural queries on profiling reports.
      5
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      MCP server analyzing Windows ETW (.etl) traces — 54 tools spanning CPU/wait stacks, image-load gaps, file/disk/mmap I/O, VirtualAlloc, network, registry, ALPC, DPC/ISR, CLR (GC/JIT/alloc/exception/contention), NT heap, and any user-mode ETW provider. Domain-neutral; works on any Windows trace.
      18
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI assistants to analyze Windows WPR/ETW traces (.etl files) using natural language queries, with features like auto-summary, CPU sampling, DPC/ISR analysis, and symbol resolution.
      14
      MIT
    • A
      license
      B
      quality
      C
      maintenance
      Enables analysis of Microsoft PIX GPU and timing captures, including event inspection, counter queries, resource views, pipeline state, and GPU timing through an MCP interface.
      67
      MIT