Experimental next generation frontend tooling.
It optimizes for memory and cold start, where running many builds (CI, agents, multi-tenant) under Vite gets expensive. OJ is meant to run real production React apps without changes to their source.
There is an SSR mode: oj dev --ssr src/entry-server.tsx in dev, and oj build --ssr for production. It streams the HTML out with renderToReadableStream instead of buffering, and the client hydrates through
the normal dev pipeline, so Fast Refresh and HMR keep working over a server-rendered page.
In dev the server modules run in a small persistent Node process (a module runner using vm.SourceTextModule) that re-evaluates only what changed, instead of rebuilding a bundle per request.
Vite/Rollup-style plugins run through a persistent Node plugin host. Drop an oj.plugins.mjs at the app root that default-exports a plugin array, or let oj read an app's vite.config.{ts,js,mjs} and pick up its plugins array directly.
A TypeScript config (including one that imports local .ts files) is loaded via Vite's own config loader when Vite is installed, or bundled with the app's esbuild otherwise. From a vite.config oj also adopts the app's base, server.port/host, define, and resolve.alias for any field its own config
leaves unset, alias entries resolve alongside tsconfig paths, in both oj dev and oj build.
Point oj at a TanStack Start app and it runs the framework directly, no source changes. It detects Start from the app itself:
oj dev web # dev server for the ./web Start app
oj build web # production buildDev serves file-based routes with streaming SSR and client hydration, keeps Fast Refresh working, runs server functions in the module runner, and applies the app's vite.config plugins (React and the app's own). oj build emits a Node server.mjs, a Cloudflare worker.mjs (nodejs_compat), content-hashed client assets, and prerendered routes. oj's own docs site (www/) is a TanStack Start app built by oj and served from a Cloudflare Worker.
Install the CLI from crates.io, then run oj in any app:
cargo install oj --locked # install the `oj` CLI
oj dev # dev server for the current app on :5199
oj build # production build into ./distOr with Nix:
nix run github:lovablelabs/oj -- devoj embeds V8, so it builds on the targets rusty_v8 ships prebuilt static libs for: Linux glibc x64/arm64, Linux musl x64/arm64, macOS x64/arm64, and Windows MSVC x64/arm64.
Generated fanout-10 React component trees, measured save-to-paint with Playwright against Vite 8.3.5 (Rolldown-based) on an Apple M5 Max, in both its default dev mode (vite) and its experimental bundled dev mode (vite-fbm). p50/p95 over 5 cold+warm restart cycles and 10 HMR edits. Measured with oj 0.2.46.
Memory is measured over the whole process tree, and reported two ways: resident set (tree RSS) and macOS physical footprint, which excludes clean and reclaimable resident pages and so bounds retained memory. Forced-GC RSS is not a column because it cannot be measured symmetrically (Node exposes an external GC handle through its inspector; embedded V8 does not) — for reference, forcing GC before measuring lowered vite's tree RSS by roughly a third in our runs, still well above its footprint. Methodology corrections owe to #202.
1,000 components (p50/p95):
| tool | cold start | warm start | reload | HMR | tree RSS | footprint |
|---|---|---|---|---|---|---|
| oj | 466/558ms | 321/327ms | 166/168ms | 57/59ms | 180MB | 136MB |
| vite | 641/648ms | 600/606ms | 167/167ms | 38/76ms | 398MB | 338MB |
| vite-fbm | 317/356ms | 317/323ms | 47/48ms | 56/59ms | 357MB | 306MB |
5,000 components (p50/p95):
| tool | cold start | warm start | reload | HMR | tree RSS | footprint |
|---|---|---|---|---|---|---|
| oj | 1280/1451ms | 1024/1038ms | 722/726ms | 61/62ms | 398MB | 146MB |
| vite | 2312/2321ms | 2133/2153ms | 710/713ms | 156/158ms | 848MB | 731MB |
| vite-fbm | 989/995ms | 998/1000ms | 177/179ms | 60/67ms | 978MB | 931MB |
10,000 components (p50/p95):
| tool | cold start | warm start | reload | HMR | tree RSS | footprint |
|---|---|---|---|---|---|---|
| oj | 2380/2395ms | 1929/1942ms | 1431/1452ms | 66/175ms | 684MB | 201MB |
| vite | 4554/4559ms | 4163/4185ms | 1383/1407ms | 152/157ms | 1414MB | 1331MB |
| vite-fbm | 1379/1381ms | 1373/1378ms | 266/273ms | 67/72ms | 1740MB | 1741MB |
Against Vite's default dev, oj starts faster at every size: cold 1.4x at 1k to 1.9x at 10k, warm 1.9x to 2.2x. Reload is level with it. Vite's experimental bundled dev starts faster than oj and reloads far faster, because it serves one bundle instead of thousands of modules. HMR matches bundled dev at every size, and beats default dev from 5k up (default dev is faster at 1k). Memory is oj's largest and most consistent win: a 136-201MB footprint against Vite's 306MB-1.7GB, from 2.5x less than default dev at 1k to 6.6x less at 10k, measured over whole process trees.
Production builds (oj build vs vite build) produce the same output on the 10,000-component app. oj build is currently slower there: 697ms against Vite's 476ms with oj's cache cleared before each build, 523ms against 479ms primed (median of 7 interleaved runs).
- oxc_transformer/examples/transformer.rs: the pipeline oj's compiler is based on
- oxc_transformer/src/jsx: JSX + ReactRefresh transform internals
- vitejs/vite-plugin-react: the Fast Refresh glue semantics oj replicates
- vite/packages/vite/src/node/server: HMR propagation,
import.meta.hotprotocol - rolldown/rolldown: plugin hook filters, the prod linker oj embeds
cargo run -p oj -- dev # dev server for ./playground on :5199
cargo run -p oj -- dev --ssr src/entry-server.tsx # streaming SSR + hydration
cargo run -p oj -- build playground # production build into playground/dist
cargo test --workspace # rust unit tests
node --test e2e/unit/*.test.mjs # js unit tests (adapter helpers)
node e2e/run.mjs # browser e2e suite
node e2e/ssr-dev.mjs # SSR dev e2e, e2e/ssr-prod.mjs for the built server
cargo test -p oj --test start # tanstack start integration (see e2e/fixtures/start-app)
node e2e/dep-optimize.mjs # dependency pre-bundle + cjs interop integration
cargo test -p oj --bin oj plugin_api_tests # build plugin API, rollupOptions, ssr, css and assetsDir options (in-process)
cargo test -p oj_server --test dev # dev server config, css, plugin hooks, ws and HMR sockets (in-process)
node e2e/config-proxy.mjs # adopt vite.config server.proxy + ignored-config warnings
node e2e/proxy-regex-context.mjs # server.proxy "^..." regex contexts (path + query), like Vite
node e2e/hmr-overlay-recovery.mjs # startup compile error: buffered overlay, reload on first update
node e2e/hmr-json-update.mjs # editing an imported .json hot-updates its boundary (?t= stamp)
node e2e/hmr-unresolved-import.mjs # missing ./import fails with an overlay; creating the file recovers
node e2e/host-binding.mjs # dev/preview --host + server.host bind all interfaces
node e2e/awkward-paths.mjs # percent-encoded filenames served, traversal contained
node bench/generate.mjs 1000 # generate a benchmark app (then npm i inside it)
node bench/run.mjs 1000 # p50/p95 benchmark vs vite
node bench/card.mjs # render bench/card.html to oj-benchmarks.pngcargo test --workspace # unit + integration suites
node e2e/run.mjs # the end-to-end suiteThe suite is organized by failure mode rather than by module ~ adversarial
input, boundary shapes, contention, injected faults, and properties that hold
for every input ~ and docs/development/testing.md describes the layers, the
fuzz targets, and the behaviours that are deliberate boundaries rather than
gaps.
MIT, see LICENSE.
OJ was created by Raphael Amorim and has since migrated to Lovable, which now maintains the project.