strobe

strobe is an open-source performance tool by Luke Parker that lets a coding agent find what is slow in an app, fix it, and prove the fix worked with measured numbers. It covers web pages, Electron apps, Bun and Node processes, and native apps on Windows, macOS and Linux. The project describes itself as measuring what users actually feel and explaining it down to the line of code, the native stack, threads, wake-ups, allocations, frames, GPU use and VRAM.

What It Is

strobe is a command-line tool written in Rust with a built-in MCP server and an agent skill, so an AI coding agent can drive it directly. It records a target on one timeline and answers with verdicts plus short handles for recordings and experiments that the agent can drill into. A core design rule is that strobe refuses measurements it cannot defend, for example a frame metric from a headless browser, and reports the refusal as the answer rather than a misleading number. It also never slows the target with CPU or network throttling.

Benchmarks as Code

Benchmarks are TypeScript files built with the @hona/strobe npm package. Each bench names its targets, such as a local URL, an installed app or a native executable, and its cases, each with a primary metric, a run step and a reset step. strobe bench compares a base and head variant, which can be git refs, folders or installed apps, and runs paired measurements until the result is decided. It can bisect a regression across commits, gate on maximum allowed slowdowns, compare screenshots, and plan how many runs a target effect size needs.

Recording and Diagnosis

strobe record captures a page, an app profile from the repo's strobe.json, or any running process by PID or name. Probes add timing, code attribution, memory leak runs, native stacks and off-CPU waits. strobe query turns a recording into hotspots, threads, frames, layout, garbage collection, memory and waits, or the difference between two recordings. strobe look shows the screen as strobe sees it, strobe open loads the full recording in Perfetto, and recordings can include video.

Pages run in an isolated Chromium, Chrome or Edge by default, with optional Firefox and WebKit engines and mobile or touch emulation. Native apps can be measured from outside through operating system events, or from inside through a native bridge that reports UI work and frames.

Pull Request Results

strobe report turns an experiment into a pull request summary with a 1200 by 630 result card in light and dark versions, or a one-line badge, and can attach it to a pull request through the GitHub CLI. The card leads with the main win, such as less CPU time or memory, and shows how many measured pairs improved.

Setup

Install the package as a dev dependency with Bun and run strobe up once per machine. Setup checks the machine, installs a helper for operating system performance events where needed, installs the strobe agent skill and registers the MCP server with OpenCode. App profiles build and launch each variant in a throwaway home directory so benchmarks never touch real user data, and strobe status reports what the machine can measure and which permission is missing.