The React Native stack I start with in 2026

Expo is the framework the React Native docs point new apps at. The forum threads about ejecting describe an older Expo. SDK 58 ships the New Architecture by default. You can still drop into Swift or Kotlin when a library is missing.

The annoying part is the upgrade. If the app sits on a very old SDK, do not jump to 58 in one pull request. Move a couple of SDKs at a time. Simon Grimm's figure for apps that still need the bare workflow is about two percent, usually a requirement Expo cannot host. I have not audited that number.

Herdr is a free, open-source Rust terminal multiplexer for coding agents. I use it when two or three features should move at once, in panes, and I want to see which agent is blocked. New harnesses ship every week. Herdr will miss one of them for a few days.

Zed is the other job. It opens fast. I use it when the fix is one line and waiting on a model is slower than typing. VS Code with Copilot, Windsurf, or Antigravity are fine if that is already where you work. The split is the point: agents in one place, hands in another.

A file in app/ is a route. It has a URL, so a deep link is the same path the screen already uses.

Expo Router v56 forked the React Navigation internals. Old @react-navigation/* imports move to expo-router/react-navigation. Expo ships a codemod for that edit. After it, you follow one release train. Native tabs and stacks, API routes, and server components live in the same folder.

The file rules take an afternoon. I would budget that afternoon before the first feature, not during it.

These are two jobs.

Expo SQLite is the database already on the phone. Drizzle keeps the schema in TypeScript and the CLI generates the SQL. Drizzle Studio is how I look at rows without building a debug screen. Migrations stay yours. If that work is the thing you wanted to avoid, this pick will annoy you.

MMKV, from Marc Rousavy, holds the small values: theme, an intro-seen flag, a token. The number people repeat is about 30 times AsyncStorage, from JSI and a memory-mapped file. I have not re-run that benchmark. It does not run in Expo Go. A production app needs a dev build anyway.

I do not start a new app's local state in Context. A theme provider is fine. A store that half the tree reads is where the extra renders show up.

Legend State is signal-based. A component reads the field it needs, and the rest of the tree stays put. Jay Meistrich's line, from his App.js Conf 2026 talk, is that state is the large hidden cost, not the list. Persist on device goes through ObservablePersistMMKV. If the app has to work offline and catch up later, there are sync plugins for Supabase and Firebase.

The cost is the model. If you think in reducers, the first week is clumsy. The community is smaller than Zustand. v3 is still moving. I would not rewrite a working Zustand app for sport. I would start the next one here.

TanStack Query is a different job. It owns HTTP: cache, retry, and the loading flag, instead of a useEffect I will forget to cancel. I add it in the same sitting as the Expo template on any app that calls a server.

FlatList is what the docs show first. I leave it on short screens.

A chat, or a feed with uneven row heights, goes to Legend List v3. It does not ask for an estimated item size. It can open on the newest message and stay pinned while new ones arrive. Same people as Legend State, which is why the two sit together.

A long uniform list, products or posts with the same row, goes to FlashList v2. Swap the component name, recycle the cells, and it runs in Expo Go. FlashList has the larger community when you get stuck. Uneven and chat-like: Legend List. Same row, many times: FlashList.

Forms are React Hook Form plus a Zod schema. The field you are typing in is the one that re-renders. The schema holds the rules, so the screen does not collect a pile of error strings.

Gesture Handler and Reanimated ship with the Expo template. Gesture Handler is the finger: tap, pan, pinch, long press. Reanimated 4 runs that motion on the UI thread. I do not keep a second animation library around just in case.

EAS is how the binary gets built, how it reaches the stores, how an update goes out, and where an API route can be hosted. Past the free tier it is a bill. I still use it. Standing up the same build, submit, and update path by hand is the slower job.

Clerk is the auth UI. The React Native components got usable over the last year. The lock-in is real. Auth is rarely the part of the app I want to be clever about, so I accept it.

Sentry goes in before the first TestFlight or internal track. Without it, the first crash is a guess. It also does session replay now. Turn that on only after you know what it records.

RevenueCat is the purchase SDK. The name people mistype is RevenueKit. It can build the paywall and create the store products. You still have to read the store rejection email yourself.

CodeRabbit reads the pull request, summarizes it, and marks the weak spots. Useful when you are the only reviewer. You still have to read the review. A summary is not a pass.

PostHog is product analytics, session replay, feature flags, and surveys, and it has a React Native SDK. Before the first event, write down which ones may leave the device.

Plenty of teams already have a server. If I am choosing, I choose Supabase: Postgres, auth, storage, edge functions. You have to be willing to write some SQL. galaxies.dev runs on it, which is Simon's reason. Mine is that I would rather learn SQL once than invent a second data model.

If SQL is a no, I would try Instant. It is closer to Firebase than to Postgres, and the client tooling is the part that has impressed people who used it for a real app, Simon included.

The third option is an Expo API route on EAS Hosting. That is where an OpenAI key lives: env file on the server, called from the route, never shipped in the binary. Right size for a secret and a thin handler. Wrong size for billing rules or a job queue.

I still start with StyleSheet. The coding agents know it, and a lot of screens do not need a styling framework.

When the app needs breakpoints, variants, or a theme without leaving that shape, I move to react-native-unistyles. You replace StyleSheet.create and keep most of the call sites. The overhead over plain StyleSheet is small.

Uniwind is the Tailwind path, from the same author. I use it when the person writing the styles already thinks in utilities. The cost is that mental model, not raw speed. If nobody on the app wants Tailwind, do not adopt it to look current.

Expo UI is the native rung. It wraps SwiftUI and Jetpack Compose components. What is wrapped feels like the platform. What is missing, you build yourself. Start at StyleSheet. Move one step when a screen forces it. Four styling libraries in one app is how a simple color change takes a day.