Perry

Write TypeScript. Ship native. Everywhere.

Perry compiles the TypeScript you already write into real machine-code executables — for macOS, Windows, Linux, iOS, Android, watchOS, and TV. No Node.js to install. No Electron to bundle. No runtime at all. Just a binary. (The same codebase can also target the web, emitted as JavaScript or WebAssembly.)

Website · Documentation · Showcase · Examples

perry compile src/main.ts -o myapp
./myapp    # a standalone native binary — ~34 KB for hello world

The language you know. The deployment you wish it had.

Millions of developers write TypeScript every day — but shipping it has always meant shipping a JavaScript engine: a Node install on every server, ~100 MB of embedded runtime per CLI, or a whole browser engine per desktop app. Perry removes the engine. SWC parses your code, LLVM compiles it to machine code, and you get what systems languages get: instant cold starts, tiny self-contained binaries, real threads — without leaving TypeScript.

	Perry	Node.js	Bun	Electron
What you ship	One native binary, from ~34 KB	Your code + a Node install	One binary embedding the JS engine	App bundle with a browser engine
Execution	Ahead-of-time machine code	JIT	JIT	JIT
Cold start	Instant — no engine to boot, no warmup	Engine boot + warmup	Engine boot + warmup	Browser boot
Native UI (no WebView)	✅ AppKit, UIKit, Android Views, Win32, GTK4	—	—	Chromium
iOS · Android · watchOS · TV	✅ from the same codebase	—	—	—
Multicore	Real OS threads, data-race-safe at compile time	worker_threads	workers	processes

_{Table describes Perry's native targets. --target web / --target wasm emit JavaScript / WebAssembly that runs in the browser rather than as a native binary.}

Performance

Ahead-of-time machine code means no engine boot and no JIT warmup — and on measured workloads, real multiples over the JS runtimes. Highlights from our open benchmark harnesses (Apple M1 Max, medians of repeated runs):

Workload	Perry	Node.js	Bun	Rust
Image convolution (4K, 5×5 Gaussian)	354 ms	1,207 ms — 3.4× slower	915 ms — 2.6× slower	392 ms
Fibonacci (recursive calls)	309 ms	987 ms — 3.2× slower	518 ms — 1.7× slower	316 ms
JSON pipeline (100 records)	39 ms	144 ms — 3.7× slower	51 ms — 1.3× slower	34 ms
JSON pipeline (500k records, 108 MB)	1,649 ms	1,010 ms — 1.6× faster	647 ms — 2.5× faster	604 ms
Object allocation (1M objects)	2 ms	8 ms — 4× slower	6 ms — 3× slower	<1 ms
Array write (10M elements)	3 ms	9 ms — 3× slower	6 ms — 2× slower	7 ms
Peak memory (JSON pipeline)	3.5 MB	36 MB — 10× more	11 MB — 3× more	1.2 MB

Look at the Rust column again: on convolution, fibonacci, and the array-write loop, TypeScript compiled with Perry runs even with — or ahead of — Rust. And Electron isn't in the table because it doesn't compete here: it ships a whole browser engine per app, where a comparable Perry app is a single-digit-MB native binary.

_{Sources: convolution & JSON from the systems-language report
— pinned to 7beb3a50ca, the last revision whose artifact contains these values; the copy at HEAD was regenerated in #7641 on a host
where the harness clock was invalid and now reads 0.0 ms in every timing cell (see this PR). Re-pin to main once it is regenerated.
Fibonacci, object allocation & array write from the polyglot sweep (Perry default mode, no fast-math).}

We publish everything, including the workloads where V8's JIT still beats us — no cherry-picked table can survive an open harness. Run it yourself: ./benchmarks/run_public_baseline.sh (methodology).

The full published suite, generated from the artifact — including the rows we lose

Generated from benchmarks/results/public-node-bun-v1.json at Perry commit b77aba63433b. Lower wall-clock median is better; every row includes complete raw samples and passed correctness checks. Medians within 5% of a peer count as a tie with it.

Benchmark	Perry	Node.js	Bun	Result	What it tests
modulo_loop	93 ms	95 ms	95 ms	tie	Integer-modulo accumulation loop (microbenchmark)
method_calls	35 ms	11 ms	8 ms	loss vs both	Class method dispatch
closure	47 ms	49 ms	49 ms	tie	Closure creation and invocation
binary_trees	13 ms	6 ms	6 ms	loss vs both	Tree allocation and traversal
string_concat	6 ms	31 ms	5 ms	mixed	String append loop
prime_sieve	6 ms	5 ms	5 ms	loss vs both	Sieve of Eratosthenes
mandelbrot	25 ms	24 ms	29 ms	mixed	Complex-number iteration
matrix_multiply	19 ms	33 ms	33 ms	win vs both	Matrix multiplication
json_roundtrip	144 ms	379 ms	219 ms	win vs both	Parse and stringify ~1 MB JSON

Why developers pick Perry

⚡ Native speed, zero warmup. LLVM-optimized machine code with escape analysis, scalar replacement, and a generational GC — see the numbers above.
📦 Binaries you can actually email. Hello world is ~34 KB: a program that only prints constants links no runtime at all. Once a program uses the runtime (loops over values, JSON, async), expect a few MB stripped — about 5.5 MB on macOS arm64 today. A real MongoDB GUI built with Perry ships as a ~7 MB app.
🔌 Your Node code mostly just works. ~97% pass rate on Node's own test suite across 53 node:* modules — real implementations of fs, http/http2, net/tls, crypto, stream, child_process, worker_threads, fetch and the web globals, plus ~50 popular npm packages (Fastify, Express, mysql2, pg, ioredis, ws, bcrypt, jsonwebtoken…). Plain JavaScript compiles too.
🖥️ One UI codebase, 11 targets. A SwiftUI-like API that compiles to real platform widgets — macOS, iOS, iPadOS, visionOS, tvOS, watchOS, Android, Wear OS, Windows, Linux, Web/WASM. Even home-screen widgets.
🧵 All your cores, safely. parallelMap, parallelFilter, and spawn on real OS threads — the compiler rejects shared mutable state, so data races don't compile.
🔋 Batteries included. Databases, WebSockets, containers, i18n, keychain, notifications, auto-update, a TUI framework — natively implemented and statically linked.

Get started in 60 seconds

# Install (macOS · Linux · Windows)
npm install -g @perryts/perry     # or: brew install perryts/perry/perry
                                  # or: winget install PerryTS.Perry

# Create and run a project
perry init my-app && cd my-app
perry run .

npm packages and ES modules work as you'd expect:

import fastify from 'fastify';

const app = fastify();
app.get('/api/users', async () => [{ id: 1, name: 'Alice' }]);
app.listen({ port: 3000 }, () => console.log('Listening on :3000'));

perry compile src/main.ts -o api && ./api    # one binary — no node_modules on the server

Same code, other platforms: --target ios, --target android, --target web… full list in the platforms guide. APT, Scoop, install script, and source builds are in the installation guide; run perry doctor to verify your setup.

Built with Perry

Real products, shipping today:

Project	What it is	Platforms
Bloom Engine	Native TypeScript game engine — Metal, DirectX 12, Vulkan, OpenGL	macOS, Windows, Linux, iOS, tvOS, Android
Mango	Native MongoDB GUI — ~7 MB binary, sub-second cold start	macOS, Windows, Linux, iOS, Android
Hone	AI-powered native code editor with terminal, Git, and LSP	macOS, Windows, Linux, iOS, Android, Web
dB Meter	Real-time sound level measurement at 60 fps	iOS, macOS, Android

More in the showcase — built something with Perry? Open a PR and add it.

Documentation

Everything else lives in the docs:

Getting started — install, hello world, project config
Language support — supported TypeScript features and limitations
Native UI · Multi-threading · Standard library
Platforms — per-platform guides from macOS to watchOS to WASM
CLI reference — commands, flags, perry.toml, privacy & telemetry
Contributing — architecture, building from source, and the release process

Community

Perry is built in the open — come say hi:

💬 Join the Discord — get help, share what you're building, and talk directly to the people building Perry
🐛 Issues — bug reports and feature requests
🚀 Showcase — apps the community is shipping

Privacy

Telemetry is opt-in: nothing leaves your machine unless you explicitly enable it in ~/.perry/config.toml, and PERRY_NO_TELEMETRY=1 (or CI=true) always wins. What can be sent is anonymous and redacted — never your source, paths, or project names. Inspect it any time with perry doctor, and see exactly what's in the payload in the privacy & telemetry docs.

License

MIT