// the take
An MCP server that gives coding agents a typed interface to real profilers and tracers — memray, nsight compute/systems, nvbench, perfetto, v8 profiles, pytest capture, rocprof, xctrace — so an agent can pull actual runtime evidence instead of guessing where the hot path is. It's for people building or using AI coding agents that need to profile CPU, GPU, native, or JS code rather than just reason about it from source.
Wide provider coverage under one contract — CPU (py-spy-style), GPU (nsight, nvbench, rocprof), memory (memray), native, and browser/JS (v8 profiles) all go through the same typed evidence model instead of one-off scripts. Explicit about epistemics: it separates observed/derived/inferred claims and states outright that profiles don't establish causality or improvement — a rare bit of honesty for a tool whose whole pitch is 'evidence.' No workspace binding by design — exact argv, absolute cwd, no searching parent directories or editing project files — avoids the usual MCP-server trust problem of an agent poking around your repo uninvited. Decent hygiene: ruff, mypy, pytest, pre-commit, CI, committed lockfile, release-please.
Provider/worker surface is large (20+ provider files, 15+ worker modules) for a 218-star project — that's a lot of third-party tool output formats (nsight, perfetto, v8) to keep parsing correctly as those tools rev their own schemas. The project's own docs/testing.md admits 'known proof gaps,' so coverage on the GPU/native paths specifically is worth checking before trusting it in anything but a toy case. Distribution is split across a Python CLI (uvx) and an npm wrapper that shells out to rewrite other tools' MCP client configs and manage version pinning/rollback — more moving parts than most MCP servers take on, and config-mutation logic is exactly the kind of code you want battle-tested before trusting it. Preservation and session-scratch model (bounded session, SHA-256 evidence bundles) sounds solid on paper but there's no indication in the README of retention limits or what happens under concurrent agent sessions hitting the same server.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →