Bench

Measure and prove performance on a real device. Bench tracks UI FPS, JS FPS, CPU, memory, and jank in real time, records runs you can save and compare, and can benchmark a matrix of variants to tell you which implementation is actually faster.

Installation

Bench relies on native peers. Install the package with Reanimated and Worklets (required), plus the performance toolkit and Nitro modules that power true native UI-thread metrics:

These ship native code, so rebuild the app with a custom dev build (expo prebuild then npx expo run:ios / run:android) — not Expo Go — and run pod install on iOS. See Reanimated's setup for its Babel plugin. Without react-native-performance-toolkit, Bench still runs in JS-only mode; adding it unlocks native UI FPS and CPU.

Once installed, Bench appears in the floating menu. On Buoy Desktop it also renders as a live HUD you can watch while you use the app.

---

What You Can Do

• Live HUD — Watch UI FPS, JS FPS, CPU, and memory update in real time as you navigate.
• Record runs — Capture a session, save it, and keep a library of recordings.
• Compare — Long-press to select two or more saved runs and see them side-by-side in the same bar-chart report batches get (duration, memory, JS FPS, CPU) to prove a change made things faster.
• Batch benchmarks — Run the same flow across several variants and get a ranked report with per-metric leaders and at-risk flags.
• Render capture — See why a run was slow: every recording also captures per-component render counts and durations (React-profiler-style), so reports show the top re-rendering components next to the FPS numbers.

---

Render Capture

With @buoy-gg/highlight-updates installed, every recording — manual or batch — also captures which components rendered, how many times, and how long they took (self time, React DevTools profiler convention). Batch reports gain a Top re-renderers section per case, single-run reports get a Render commits block, and the MCP run_benchmark_batch ranking includes the heaviest components per case — so instead of "case B dropped to 41 JS FPS" you get "case B dropped to 41 JS FPS because ProductList rendered 47× costing 312ms".

A few things to know:

• Requires a dev build (the React DevTools hook and profiling timers aren't present in release builds). Without them the run simply has no render data — nothing breaks.
• Capture walks committed fibers on the JS thread, adding a small overhead per commit. It's applied uniformly to every case in a batch, so relative comparisons stay fair — but for absolute FPS measurements you can turn it off with the Capture renders toggle in Automate settings, or captureRenders: false on run_benchmark_batch.
• Buoy's own devtools UI (the HUD, floating menu, etc.) is excluded from results automatically.

---

Optimize with your AI

Bench powers Buoy's AI performance workflow. The MCP server ships a buoy-optimize wizard skill that turns the whole tune → measure → repeat loop into an almost hands-off process. Point your assistant at a new feature you're building — or an existing screen that's slow on device — and it will:

• Navigate to the screen and benchmark a matrix of implementation variants with run_benchmark_batch
• Read the ranked comparison (FPS, CPU, memory, jank) from the real device, pick the winner, and apply the next optimization
• Loop — measure, change, re-measure — until the numbers stop improving

Because it measures on-device instead of guessing, it collapses what is normally days or weeks of AI back-and-forth into minutes. In one real pass, a Skia LED display went from 28 lights stuttering under advanced effects to over 12,000 lights with no lag.

The only steps that stay manual are the ones a human has to eyeball — mainly confirming the result still renders correctly. If you're on Skia or other GPU-drawn UI, plan to glance at the screen between passes: the wizard drives the metrics, you verify it still looks right. Say "buoy optimize" in your editor to start.

---

What's Next

• AI / MCP Server — Automate benchmarks with an AI agent
• Render Highlighter — Find the re-renders hurting performance
• Events Timeline — See performance in context with everything else