Metal GPU Trace MCP server

metal-gpu-trace

Measure how a Wine or macOS program uses the GPU on an Apple silicon Mac, from the command line or through an AI coding agent: GPU time per frame, frame pacing, FPS and memory, recorded with Instruments' Metal System Trace; whether a translation layer (DXMT, Mullion, MoltenVK) splits the work into too many short encoders; Metal .gputrace captures read without Xcode; and Apple's published facts on each GPU family to read the numbers against. Python 3.9+ standard library, zsh, and Xcode (for xctrace). MIT license. Español.

Architecture, decisions and guides: docs/README.md. Sending a change: CONTRIBUTING.md.

File	What it does
src/metal_gpu_trace/perf_record.sh	Records a program for N seconds with Metal System Trace and writes <label>_<date>.json (GPU ms per frame, frame intervals, FPS, Metal and resident memory) plus per-frame and memory CSVs. Deletes the .trace afterwards unless KEEP_TRACE=1.
src/metal_gpu_trace/perf_report.py	Summarizes an existing .trace the same way.
src/metal_gpu_trace/gputrace_read.py	Reads a Metal .gputrace capture without Xcode: command buffers, encoders, draws, debug groups. Call names marked probable are not confirmed.
src/metal_gpu_trace/data/apple-gpu-families.json	Apple7 to Apple10: chips, what each family adds, limits that differ, with Apple's sources.
src/metal_gpu_trace/metal_trace_mcp.py	An MCP server over all of the above.
tests/	Self-check without Instruments: python3 -I tests/run.py prints ok; pytest tests runs the same tests.

Recording

PERF_OUT=~/perf-runs zsh src/metal_gpu_trace/perf_record.sh game-d3d12 30 /path/to/wine ~/.wine /path/to/game game.exe
PERF_OUT=~/perf-runs zsh src/metal_gpu_trace/perf_record.sh app-metal 30 native - /path/to/app ./app

Arguments: label, seconds, a Wine directory holding bin/wine (or native), the Wine prefix, the working directory, the program and its arguments. Without PERF_OUT, runs go to ./perf-runs. Reading the trace afterwards takes about a minute on an M1 Pro (87 s for a 10 s recording); through the MCP server the whole call times out after 10 s per recorded second plus 120 s, and an overrun recording leaves no files behind. One recording runs at a time; keep other GPU work off the Mac while it records.

Under Wine, give it 30 s: Wine's start-up takes part of the recording. A recording with under 100 frames is too short to compare.

Disk use

Each recording keeps a JSON summary and three CSVs, about 2 MB for 10 s (most of it the Metal memory samples). The .trace (tens to hundreds of MB) is deleted unless KEEP_TRACE=1. KEEP_RUNS=N keeps only the newest N recordings. PERF_MEMORY=0 skips the Metal memory table, the slowest to read: the summary then has no metal_memory_mib and there is no memory CSV (41 % faster to summarize a 626 MB trace). The MCP tools perf_record and perf_baseline_record offer the same as memory=false.

MCP server

Install

The server runs with uvx from PyPI, with nothing to clone (brew install uv installs it). Recording also needs Xcode installed (xcrun xctrace version prints a version).

uvx metal-gpu-trace

This is the command each client below runs. Desktop apps do not see Homebrew's PATH: give them the full path that which uvx prints (/opt/homebrew/bin/uvx on Apple silicon). uvx keeps the version it first downloaded; run the command once with --refresh after uvx to get the latest. uvx metal-gpu-trace@0.4.0 pins a version.

The server reads these environment variables:

Variable	What it sets
METAL_TRACE_OUT	Where recordings go. Give an absolute path: the default, ./perf-runs, depends on the client's working directory.
METAL_TRACE_WINE	The Wine directory (holding bin/wine) that perf_record uses when the call names none.
WINEPREFIX	The Wine prefix used when the call names none; ~/.wine otherwise.
METAL_TRACE_READ	More directories the tools may read, separated by :.
METAL_TRACE_LOG	off silences the one-line stderr log per call.
KEEP_TRACE	1 keeps each recording's .trace, for trace_tables, trace_query or Instruments.
KEEP_RUNS	Keeps only the newest N recordings.
PERF_ENV	NAME=VALUE pairs, separated by spaces, set for the recorded program.
PERF_INSTRUMENTS	Instruments added to the Metal System Trace template, separated by ; (xcrun xctrace list instruments names them); their tables need the .trace, so pair it with KEEP_TRACE=1.

Claude Code

claude mcp add --scope user metal-trace -e METAL_TRACE_OUT=$HOME/perf-runs -e METAL_TRACE_WINE=/path/to/wine \
  -- uvx metal-gpu-trace
claude mcp list

claude mcp list shows metal-trace as connected. --scope user makes it available in every project; leave it out to add it to the current project only.

Windows

On Windows the server records native programs (src/metal_gpu_trace/perf_record_windows.py); no Wine and no Metal. METAL_TRACE_WINE is not needed: perf_record and perf_attach use native. METAL_TRACE_READ is separated by ;.

Frames come from Intel's PresentMon console application (MIT): one row per Present() of the process, with MsBetweenPresents and the GPU time MsGPUTime. Download PresentMon-<version>-x64.exe from its releases, check its SHA-256 against the one GitHub shows, and call the presentmon_install tool, or put it in %LOCALAPPDATA%\metal-gpu-trace or on PATH, or point PRESENTMON_EXE at it. The user must be an administrator or in the "Performance Log Users" group. Without PresentMon the summary falls back to DWM composition intervals (gpu.source is dwm-composition): they cover the whole desktop, miss games in exclusive full screen, and carry no GPU time. With neither a frame log, PresentMon nor DWM samples, a recording fails; it never invents frames.

What cannot be measured is null, not zero: the 1% low needs 100 frames and the 0.1% low 1000; stutter rate and jitter need two frames; thermal_throttling_detected needs GPU times; a drawstress log holds one mean per line, so its summary (gpu.frame_times) has no frame-level percentiles, stutter or jitter. DWM keeps only intervals between consecutive composed frames and drops a jump of several frames. perf_diagnose calls a large working set with uneven frame times MEMORY_PRESSURE_SUSPECTED: a suspicion, since memory bandwidth is not measured.

winget install --id=astral-sh.uv -e
claude mcp add --scope user metal-trace -e METAL_TRACE_OUT=%USERPROFILE%\perf-runs -- uvx metal-gpu-trace
claude mcp list

To run from a clone instead of PyPI (needs Python 3.9 or newer):

claude mcp add --scope user metal-trace -e METAL_TRACE_OUT=%USERPROFILE%\perf-runs -- python C:\path\to\metal-gpu-trace\src\metal_gpu_trace\metal_trace_mcp.py

In JSON clients use "command": "uvx" (or the full path where uvx prints) and Windows paths with doubled backslashes. Claude Desktop reads %APPDATA%\Claude\claude_desktop_config.json. Run the Windows tests with python -m pytest tests/test_perf_record_windows.py.

Claude Desktop, Cursor and other JSON clients

These apps do not expand ~ or $HOME: write absolute paths.

{
  "mcpServers": {
    "metal-trace": {
      "command": "/opt/homebrew/bin/uvx",
      "args": ["metal-gpu-trace"],
      "env": {
        "METAL_TRACE_OUT": "/absolute/path/to/perf-runs",
        "METAL_TRACE_WINE": "/absolute/path/to/wine"
      }
    }
  }
}

Client	File
Claude Desktop	~/Library/Application Support/Claude/claude_desktop_config.json (Settings, Developer, Edit Config); restart the app
Cursor	~/.cursor/mcp.json
VS Code	.vscode/mcp.json in the project; the top key is servers instead of mcpServers, and the server adds "type": "stdio"

Codex CLI

In ~/.codex/config.toml:

[mcp_servers.metal-trace]
command = "/opt/homebrew/bin/uvx"
args = ["metal-gpu-trace"]
env = { METAL_TRACE_OUT = "/absolute/path/to/perf-runs", METAL_TRACE_WINE = "/absolute/path/to/wine" }

Tools

Tool	Reads or records
perf_record	Records a program and returns its summary; when it ends it stops every program of its Wine prefix, or every process named like a native program. counters=true also records Metal GPU Counters: gpu_counters (the limiters, utilization, occupancy and memory bandwidth of the GPU, with the main limiter) in the summary, at about 40 s of read-back per recorded second; not under Windows
perf_attach	Records a program that is already running, by pid or name, and leaves it running
perf_runs	Lists recordings, newest first, one page at a time, whether each kept its .trace, and the translation layer it ran on (a fingerprint of its Direct3D DLLs, and its Wine)
perf_compare	Compares two recordings: headline numbers, encoders, GPU idle time and each encoder label's GPU ms per frame, largest change first; flags those under 100 frames. markdown=true adds the numbers as Markdown tables for a PR comment; skip_s leaves out the first seconds of each recording (start-up) and recomputes the GPU numbers, hitches and Metal memory without them. perf_gate and perf_panel take skip_s too
perf_baseline	Saves several equal recordings as a named reference and returns the range each number spans across them and spread_pct, how much equal runs differ (the noise a change must beat); alone, lists the references. skip_s saves each run without its first seconds and perf_check applies the same cut to the run it judges
perf_baseline_record	Records the same program count times (default 3) with perf_record's arguments and saves the runs as a reference in one call. skip_s as in perf_baseline
perf_cold_warm	Records the same program twice in a row and says how many more shader-compile hitches the first run has than the second (first_seen_hitches), with each run's hitches, hitches per minute and graphics compiler time; the Metal shader cache is not cleared between them
perf_ab	Records the same program with two environments (env_a, env_b: NAME=VALUE pairs) in the order A B B A and says per metric (FPS, 1 % low, GPU ms, stutter) whether B is better, worse or the same: better or worse only when every B value is outside A's range
perf_sweep	Records the same program once per value of one variable (dim, size, example, api or env:NAME), 2 to 8 values, rounds times (every other round in reverse order), and returns per value the median FPS, 1 % low, GPU ms per frame and stutter; with 3 or more numeric values it fits GPU ms per frame on the value (fixed ms and ms per draw for dim) with its r2
perf_check	Judges a recording against a reference: regressed when FPS falls under the range by more than margin_pct (default 3), GPU ms per frame rises over it, or encoders or command buffers per frame exceed it; improved or same otherwise, with the reasons and the full comparison
perf_gate	Checks a recording against budgets you set (FPS and 1% low floors; stutter, p95 GPU ms, hitch and Metal MiB ceilings) and returns pass or fail with each limit and value; a missing number fails
perf_history	Follows FPS, 1% low, GPU ms, stutter, hitches and Metal MiB across the runs whose name starts with a prefix, oldest first, with the change from first to last
perf_panel	Writes <name>.panel.html: a page with GPU time per frame and its hitches, the numbers, Metal memory and, for two runs, A against B. It has no server and loads nothing from the network; the viewer hides or reorders widgets and exports the layout as JSON. With --http, /panel#token=...&runs=a,b shows it live, refreshed every 10 s. layout (tab ids in order) saves panel.json in the runs directory as the default tabs of every panel, the live one included; the GPU counters tab shows when a run has them
perf_export_trace	Writes a recording's GPU frames as <run>.chrome-trace, one slice per frame, for ui.perfetto.dev or chrome://tracing
perf_diagnose	Says whether a recording splits its GPU work into too many short encoders, as translation layers do, with encoders and command buffers per frame, the labels with the most GPU time and the GPU's idle time
trace_tables	Lists the tables of a kept .trace, named by its run or its path
trace_count	Counts the rows of several tables of a kept .trace, in all and per process; by default the tables that matter for a translation layer (GPU intervals, encoders, compiler spills, compiled shaders, allocations, counters, Time Profiler), so an empty one shows before reading it
trace_query	One page of rows of one table, optionally one process
capture_summary	Summarizes a .gputrace; lists its 100 most frequent calls out of calls_total
capture_records	One page of a .gputrace's records
gpu_family	Apple's facts for a family or chip (Apple7, M1 Pro, M3 Max)
presentmon_install	Windows only: downloads PresentMon, checks its pinned SHA-256 and keeps it in %LOCALAPPDATA%\metal-gpu-trace, where perf_record and perf_attach find it; Claude Code asks before every call
workload_list	Lists the test programs workload_fetch can download; for bgfx-examples, example_traits gives each example's class and its encoders per frame in native Metal
workload_fetch	Downloads a test program once, checks its SHA-256 and keeps it in ~/Library/Caches/metal-gpu-trace; Claude Code asks before every call

perf_record takes workload, api (mtl, mtl-x64, vk, d3d11 or d3d12), dim, size and example instead of program, to record a fetched test program (workload_list says what each has). mtl-x64 is native Metal built for x86_64 and run under Rosetta, like Wine's programs; size ("1920x1080") sets drawstress's window. The first workload is bgfx's 17-drawstress (BSD-2, 9.3 MB), built from source by .github/workflows/workloads.yml: a grid of cubes, dim³ draws per frame, with its shaders inside the program. The compare-layer prompt (/metal-trace:compare-layer drawstress d3d12 in Claude Code) records it in native Metal and under the Wine in METAL_TRACE_WINE, then runs perf_diagnose and perf_compare.

The second workload, bgfx-examples (55 MiB), holds thirty bgfx examples, each a different load (many cubes, textures, render-target chains, a G-buffer, compute, reflective shadow maps, post-processing, draws the GPU decides, text and interface, shadows, transparency, GPU read-back, virtual textures, upscaling, texture formats, occlusion queries, tessellation, multiple windows, off-screen rendering; the last twenty-one measured in native Metal and Wine D3D12 only, example_traits in workload_list gives each one's class and encoders per frame in both, and 40-svt writes a 370 MB tile cache to temp/ in the workload's folder on its first run): perf_record(workload="bgfx-examples", api="d3d12", example="21-deferred"). Two macOS levers change the Mac under a recording: memory_free_pct (20 to 60) holds the whole Mac at that percentage of RAM free (memory_pressure), and background runs the program under taskpolicy -b, a hard throttle. A small GPU-bound program does not feel the first (drawstress stayed at 728 FPS and its resident memory fell); the second slows it several times over. Repeat a run that uses either.

The third workload, godot, is a lit 3D scene a real engine builds from code (Godot 4.7.2, Forward+): 400 meshes, 8 lights, a shadowed sun, api="mtl" for Godot's Metal driver and api="d3d12" under Wine for its Direct3D 12 driver, with args="-- meshes=800 lights=16" to change it. workload_fetch("godot") keeps the 10 KB project from this repository's release and downloads Godot's official macOS (171 MB) and Windows (86 MB) builds from godotengine/godot-builds, checking the SHA-512 this package pins; it needs about 257 MB and some seconds. The scene has no imported assets, so Godot does not import anything. Godot compiles its pipelines on every start, which makes it a case for perf_cold_warm: its shader-compile hitches came back unchanged in the second run.

Reading the numbers across runs

• A summary says where it ran and what the GPU did: machine (chip, cores, RAM, macOS, screen_locked), gpu_activity (utilization of the device, renderer and tiler, split scenes, GPU recoveries: it covers the whole GPU, not only the program), resource_allocations (allocation events per frame, by type; a running total, not memory in use) and command_buffer_errors.
• A locked Mac session leaves native Metal windows unpaced: 3 000 to 4 000 FPS against 120 with vsync, on the same examples. Record with the Mac unlocked; perf_compare returns screen when one run was recorded locked and the other not.
• GPU ms per frame follow the GPU's performance state: a run paced at 120 FPS sits in Minimum and takes more ms for the same work (21-deferred: 1.75 ms paced, 0.27 ms unpaced). perf_compare returns gpu_state when two runs sat in different states, and perf_diagnose returns display_synced_hz when the FPS sits on a display rate. Compare runs paced alike: native Metal and D3D12 under Wine with Mullion, both at 120 FPS, do compare.
• With counters=true, each listed hitch also has top_limiter: the Limiter counter with the highest mean over the interval before it (window_ms, read in whole 10 ms bins; null when no sample falls in it), which can differ from the whole run's main limiter.

api=vk runs the same program through Vulkan on MoltenVK. MoltenVK is not shipped: set METAL_TRACE_MOLTENVK to a directory (no spaces) holding an arm64 libMoltenVK.dylib, such as the one in a Vulkan SDK or brew install molten-vk. The run's layer names MoltenVK with the library's SHA-256, so perf_compare against a native Metal run says it compares two layers. In one measurement at dim=20, 10 s each: native Metal 2 encoders per frame; MoltenVK 13, 91 % of them short, and the GPU idle 89 % of the time (perf_diagnose: too_many_short_encoders true). One pair of runs, not a benchmark.

api=vk is native only. drawstress.exe --vk under Wine needs a Wine built with Vulkan: the Wine of DXMT and Apple's Game Porting Toolkit were tried and both log Wine was built without Vulkan support, so the recording holds no GPU frames. CrossOver's Wine has Vulkan but runs only inside one of its bottles, which was not tried.

The same facts are MCP resources: metal-trace://gpu-families (all) and metal-trace://gpu-families/apple7 to apple10. Each saved run is a resource too: metal-trace://runs/<run> reads its summary. They restate Apple's Metal Feature Set Tables of 21 May 2026, MTLGPUFamily and the M3 tech talk; check a device with supportsFamily.

How it follows the MCP specification (2025-11-25 and 2026-07-28) and Claude Code's limits:

• stdout carries only MCP messages; each tool call logs one line to stderr (METAL_TRACE_LOG=off silences it), with no arguments or paths.
• Lists come in pages (offset, limit, default 50, at most 500) with total and next_offset; a result over 40 000 characters is refused with a hint to ask for a smaller page.
• Files are read only inside METAL_TRACE_OUT and the directories in METAL_TRACE_READ (separated by :).
• Failures come back as isError results that say what to do, without stack traces or local paths. Summaries cut absolute paths to file names, so they can be shared.
• Every xctrace call has a timeout.

HTTP

METAL_TRACE_TOKEN=<secret of 16+ characters> metal-gpu-trace --http [--port N] serves the same tools at POST http://127.0.0.1:8765/mcp (Streamable HTTP, one JSON reply per request), for a client on the same Mac that cannot start a stdio server. Every request needs Authorization: Bearer <token>. It listens on 127.0.0.1 only, refuses a request whose Host or Origin is not localhost or 127.0.0.1 (DNS rebinding), and does not start without the token. It speaks both MCP eras: a client that sends initialize gets the behaviour it always had; one that sends MCP-Protocol-Version: 2026-07-28 is stateless and must also send Mcp-Method (and Mcp-Name for tools/call, prompts/get and resources/read), which the server checks against the body (400 with -32020 when they differ). A modern client cancels a running call by closing the connection; an initialize client sends notifications/cancelled from another connection. There is no progress stream over HTTP, and no remote access: reach another Mac through an SSH tunnel (ssh -L 8765:127.0.0.1:8765 mac), not by opening the port. Why two eras: docs/adr/0002-dual-era-mcp.md.

Windows tests on a real machine

GitHub's windows-latest has no GPU and no desktop session, so PresentMon and DWM recordings are only tested on a real Windows machine. windows/winbridge-run-tests.ps1 updates a checkout there and runs tests/test_perf_record_windows.py; it is written as a WinBridge task (fixed name, script pinned by SHA-256, no arguments). To register it, copy it to the laptop, set $repo to the checkout's path, take its hash there with Get-FileHash <file> -Algorithm SHA256 (Git may have changed the line endings, so the hash of the copy in this repo is not the one to use), and add a [tasks.<name>] table with script and sha256 to the agent's TOML.

CI gate

metal-gpu-trace check RUN BASELINE [--margin-pct N] runs perf_check without a client and prints one JSON line (verdict, reasons, hints). Exit code 0: same or improved; 1: regressed; 2: could not be judged (no such run or reference, or under 100 frames). Runs are read from METAL_TRACE_OUT. Recording needs Instruments, so the job runs on a self-hosted macOS runner:

jobs:
  perf:
    runs-on: [self-hosted, macOS, ARM64]
    steps:
      - run: PERF_OUT=~/perf-runs zsh path/to/metal-gpu-trace/src/metal_gpu_trace/perf_record.sh pr 30 native - . ./app
      - run: METAL_TRACE_OUT=~/perf-runs metal-gpu-trace check "$(ls -t ~/perf-runs/pr_*.json | head -1 | xargs basename -s .json)" main-ref

Credits

• The record layout of a .gputrace command stream follows tmc/gputrace (MIT); no code is copied. The call ids were matched against Xcode's command list on macOS 27.0.1.
• GPU family facts: Apple's documents linked above, restated, not quoted.