Compiler

The compiler (@point0/compiler) is the source transform that runs on every file as the engine builds or serves your app. On a single file, it:

• strips point method bodies per side — your loader and your component live in one file, and each side only ships the parts it needs (the headline feature, below);
• shakes env.* guards and your process.env.X constants down to a literal, then deletes the dead branch and its now-unused imports;
• swaps forbidden imports for virtual modules (server-only / client-only markers, mock/deny) so a leak fails the build;
• finds the points themselves and powers the short .lets notation by rewriting it into the explicit form;
• compiles .md / .mdx / .mdc through MDX and routes static assets (?url, ?react, …) before its other passes;
• runs your babel plugins and keeps the HMR boundary alive so any mix of points hot-reloads;
• caches every result on disk, so after the first run it only recompiles files that actually changed.

It's the same compiler in every form — a Bun plugin, a Vite plugin, and a Babel plugin wrap one class; see One compiler, any bundler. You rarely touch it directly — the point0 CLI and those plugins drive it, and it's on by default for every side.

See it work: point0 compile

The point0 compile command prints the compiled output for a single file:

point0 compile src/lib/prisma.ts --server   # how the file looks in the server bundle
point0 compile src/pages/home.tsx --client  # ...and in the client bundle

The same source comes out different per side — that difference is the whole point of the compiler. Useful flags:

point0 compile <file> --client          # compile for the client side
point0 compile <file> --server          # compile for the server side
point0 compile <file> --scope <scope>   # points scope (inferred from side if omitted)
point0 compile <file> --no-babel        # skip your babel plugins, show point0-only transforms
point0 compile <file> --no-hmr          # don't inject the HMR decoy (see below)

Full flag list is in the reference at the bottom.

What gets stripped from a point

This is the headline: you write the loader and the component in the same file, and each side's bundle only carries the methods it actually runs. The compiler reads the point chain and removes the bodies that don't belong to the side it's compiling — keeping the chain shape (so the call still type-checks and runs), just emptying the callbacks that were stripped.

Take one page with a server loader and a client component:

// src/pages/idea.tsx
import { prisma } from '@/lib/prisma'
import { SomethingForClient } from '@/components/something-for-client'

export const ideaPage = root.lets
  .page('/ideas/:id')
  .loader(async ({ params }) => {
    const idea = await prisma.idea.findUniqueOrThrow({
      where: { id: params.id },
    })
    return { idea }
  })
  .page(({ data: { idea } }) => (
    <div>
      <h1>{idea.title}</h1>
      <SomethingForClient />
    </div>
  ))

On the client, the server-only methods are cut from the client bundle — their bodies and the imports they use are removed, so they never ship to the browser. That's the loader and everything else that's server-only (.ctx, .input, .middleware, .headers, .cookies, .body, .response, .description, .openapi). The client still knows the point's name and route, so it can still call the loader over the network — it just doesn't carry its code (and prisma, now unused, is pruned with it):

// point0 compile src/pages/idea.tsx --client
import { SomethingForClient } from '@/components/something-for-client'

export const ideaPage = root.lets
  .page('/ideas/:id')
  .loader() // body gone — prisma never reaches the browser
  .page(({ data: { idea } }) => (
    <div>
      <h1>{idea.title}</h1>
      <SomethingForClient />
    </div>
  ))

On the server, what happens to the .page body depends on ssr.

ssr: true vs ssr: false

ssr is an engine option (per client, or a default at the top):

export const engine = Engine.create({
  // ...
  clients: [
    {
      // ...
      ssr: true, // server-render the first load; default is false
    },
  ],
})

The render methods (.page, .layout, .component, .provider, .with, .wrapper, .mapper, .head, and the .loading / .error family) are cut from the server bundle when ssr: false — bodies and the imports they use removed from the server build. The client build always keeps them.

• ssr: true — nothing is cut: the server renders the first page load, so it keeps those render-method bodies, and the client ships them too for client-side navigation after that first render.
• ssr: false — those render bodies (and their now-unused imports) are removed from the server bundle; the server keeps only the loader (to answer data requests) and the rest of the server-only chain. This is exactly the compile output of the page above when ssr is off — .page() empties on the server, .loader() empties on the client.

Client-only methods are cut from the server bundle: .clientLoader, .clientInput, .clientOn, .clientOnPrefetchPage, and the navigation/prefetch triggers (.prefetchPageOnLinkHover, .prefetchPagePolicy, .scrollRestore, …) have their bodies and imports removed regardless of ssr — that code never ships to the server (it only runs in the browser). Server-only methods (.loader, .serverOn, .serverOnPrefetchPage, …) are cut the other way, from the client bundle. .onPrefetchPage is the one prefetch hook kept in both bundles (it runs on the server once before the first render and on the client during prefetch).

Every method's own page repeats this in its terms: which bundle the method is cut from, and how ssr changes it. Read it there for the exact rule per method.

.clientOnly() — turn off SSR for the rest of one point

.clientOnly() makes a single point behave as if ssr: false for that point only — from that call onward in the chain, the render methods (.page, .with, .component, …) are cut from the server bundle — their bodies and the imports they use removed — even when the app is otherwise ssr: true. Use it for UI that genuinely can't render on the server (it touches window, a client-only library, etc.). You can pass a server-side fallback to show while the client takes over, or call it with no argument at all:

export const dashboardPage = root.lets
  .page('/dashboard')
  .loader(async () => ({ stats: await loadStats() }))
  .clientOnly(() => <Skeleton />) // server renders the skeleton…
  .page(({ data }) => <Charts stats={data.stats} />) // …this body is stripped from the server

export const WidgetComponent = root.lets
  .component()
  .clientOnly() // no fallback — server renders nothing for this point
  .component(() => <BrowserOnlyWidget />)

Everything before .clientOnly() (here the .loader) still runs on the server as usual; only the chain after it goes client-only. See SSR for when you'd reach for this.

env shaking and virtual modules

Method stripping handles the point chain. For plain code outside the chain — secrets, server libraries, branches that should only exist on one side — the compiler offers two more tools.

env.* shaking and dead-code elimination

When you guard code with the env helpers from @point0/core, the compiler replaces the guard with a literal true / false for the side it's compiling, then deletes the dead branch:

import { env } from '@point0/core'

const value = env.side.is.server ? readSecret() : 'public'
// compiled for client → const value = false ? readSecret() : 'public'
//                     → const value = 'public'   (readSecret import pruned too)

env.side.define makes this explicit — the matching side keeps the value, the other gets undefined and the code is pruned:

// examples/better-auth/src/lib/auth/server.ts
export const authServer = env.side.define.unsafe.server(
  betterAuth({ database: prismaAdapter(prisma, { provider: 'sqlite' }) }),
)
// on the client this whole expression becomes `undefined` — better-auth and
// prisma never reach the browser bundle

The same shaking covers env.mode.* (production/development/test), env.build.was, env.runtime.*, env.os.*, env.scope.*, env.ssr.active/phase/target (client-side constants — SSR is never underway in the browser), and env.vars.*, plus process.env.X / import.meta.env.X against your configured constants. Each resolves to a literal at compile time; the now-dead branches and their unused imports are removed by a dead-code-elimination pass. The full list lives on the env page.

env.runtime.is.<X> covers browser, reactNative, nodejs, bun, deno, and worker; env.os.is.<X> covers ios, android, linux, mac, and windows. The compiler shakes these only when you set the matching runtime / os option (below).

The compiler also handles <ClientOnly>: on the server, its children are replaced with null, so client-only UI never renders server-side.

Import stripping → virtual modules

The other half is the importer: instead of (or in addition to) env guards, you can forbid a whole import from one side. The compiler swaps the import for a virtual module so the real module never enters that bundle.

The server-only / client-only markers are the common case:

// examples/basic/src/lib/prisma.ts
import '@point0/core/server-only' // importing this from client code throws

export const prisma = new PrismaClient(/* ... */)

When the compiler builds the client side and finds an import that reaches @point0/core/server-only (directly or through a denied import), it replaces it with a virtual module that throws at import time — so a leak fails the build, not in production. The marker files themselves are empty (export {}); all the work is the compiler's.

You can also configure this in the engine, matching by import target:

// examples/expo/src/engine.ts
compiler: {
  importer: {
    mock: ['react-native', 'expo-router'], // replace with a no-op mock module
  },
}

• mock swaps the import for a recursive no-op proxy (createMock()) — every access returns another mock, so the import resolves but does nothing.
• deny swaps it for a throwing virtual module (like the markers).
• The deny error includes a point0 trace ... command you can run to see the import path that pulled the forbidden module in.

Full configuration — patterns, cold, onDeny — is on the importer page.

The HMR decoy

Bun's native bundler and Vite only enable React Fast Refresh for modules that look like they export a React component. A point file exports points, not components, so without help every edit would trigger a full page reload.

The compiler fixes this by appending a _tail decoy to the last point in each chain in the file. Put a mutation and the page that uses it side by side:

// you write — one file, a mutation and a page:
export const ideaUpdateMutation = root.lets
  .mutation()
  .input(z.object({ id: z.string(), title: z.string() }))
  .loader(async ({ input }) =>
    prisma.idea.update({ where: { id: input.id }, data: input }),
  )
  .mutation()

export const ideaPage = root.lets.page('/ideas/:id').page(() => <h1>Idea</h1>)

// the compiler appends a decoy to each (in dev):
export const ideaUpdateMutation = root.lets /* ... */
  .mutation()
  ._tail(() => null)
export const ideaPage = root.lets /* ... */
  .page(() => <h1>Idea</h1>)
  ._tail(() => null)

Each export is the function returned from _tail(() => null), so both Bun and Vite treat both exports as ordinary React components and wire up Fast Refresh for the module — even though one is really a mutation. You never access ideaUpdateMutation directly, only its methods, and they're all still in place; the real runtime export is decided by Point0._tail in @point0/core (a mountable returns its mount component, everything else returns the decoy). That's why a file can export any mix of pages, queries, and mutations and still hot reload — see Dev. The decoy is on by default for the client and off for the server; point0 compile --no-hmr turns it off.

User babel plugins

Your own babel plugins and presets run inside the compiler, after its own transforms. Set them under compiler.babel in the engine config — at the top for a default, or per client:

// examples/basic/src/engine.ts
export const engine = Engine.create({
  // ...
  clients: [
    {
      // ...
      compiler: {
        babel: ['babel-plugin-react-compiler'], // array form = plugins
        // or the object form for presets too:
        // babel: { plugins: ['babel-plugin-react-compiler'], presets: [...] },
      },
    },
  ],
})

The compiler has special handling for React Compiler. React Compiler only memoizes things that look like top-level components, but point logic lives in chain callbacks (.page, .layout, .component, .provider, .wrapper, .with) — so the compiler injects a "use memo" directive into those callbacks to get them memoized too. The behavior keys off React Compiler's own compilationMode option, which you pass to the plugin the usual way:

compiler: {
  babel: [['babel-plugin-react-compiler', { compilationMode: 'infer' }]],
}

• unset / 'infer' / 'syntax' → keep the plugin and inject "use memo".
• 'all' → keep the plugin, no directive (it compiles everything anyway).
• 'point0' → strip React Compiler from babel (you run it elsewhere, e.g. via Vite) but still inject "use memo".

MDX, Markdown, and assets

The compiler also claims .md, .mdx, and .mdc files — it compiles them to JS (via @mdx-js/mdx) before its other passes run, so env shaking and <ClientOnly> work inside MDX too. Configure remark/rehype/recma plugins through compiler.markdown. Details on MDX.

Static asset imports (import logo from './logo.png', ?url, ?file, ?text, ?react for SVGR) also ride through the compiler plugin. See Assets for modes and config.

Caching

The compiler caches every result on disk, keyed on the file's contents (via its mtime) and the compiler's own settings. The first run pays full price; after that a file recompiles only when it changes — or when the settings that produced its cache entry change. Cache is on by default (compiler.cache); clear it with:

point0 prune

One compiler, any bundler

Every integration is a thin wrapper around one class: a Bun plugin, a Vite plugin, and a Babel plugin all run the exact same transform. That lets a Point0 codebase target anything with a bundler — a web client on Bun or Vite, an Expo app, an Electron build — without a different compiler per target. You almost never construct it yourself; the engine wires the right plugin in for you:

Each plugin accepts either a CompilerOptions object (it builds a Compiler for you) or an existing Compiler instance. These are subpath exports — they are not in the package's main barrel.

A file is always compiled for one specific target — a fixed side, scope, and mode — so the same source runs through fresh per side and yields two different outputs (which is why --client and --server differ above). The compiler does not strip TypeScript syntax itself; it parses TS, transforms it, and emits TS, leaving the bundler (Bun / esbuild / Vite) to handle types.

Choosing Bun or Vite for the client is an engine config decision, not a compiler one — the same compiler runs on either path. See Bun or Vite.

The .lets sugar transform

The compiler powers the short .lets notation by rewriting it into the explicit form. The short form only works with the compiler on; the explicit form works anywhere and is identically typed.

// what you write (short form):
export const ideaPage = root.lets.page('/ideas/:id')./* ... */.page()

// what the compiler emits (explicit form — runtime-identical):
export const ideaPage = root.lets('page', 'idea', '/ideas/:id')./* ... */.page()

It reads two things off your code: the point type (.lets.page(...) → 'page', .lets.layout(...) → 'layout', and so on) and the point name, taken from the variable name with the type suffix stripped (ideaPage → idea). The extension points work the same way: .lets.base(...) → 'base' and .lets.plugin(...) → 'plugin', with the name inferred from the variable (myPlugin → my, authBase → auth). The full name-inference rules and edge cases are on the points page.

The compiler only desugars .lets.<type>() when the chain's base traces back to a Point0 imported from @point0/core — a look-alike on an unrelated object is left alone:

import { fake } from './somewhere'
export const x = fake.lets.page('/ignore') // NOT a point — left untouched

Reference

point0 compile flags

point0 compile <file> — print the compiled output for one file.

-h is rebound to --hmr here; use --help for help.

Omit --scope and the compiler infers it from the side: the server side uses the server scope, the client side uses its one client scope. With more than one client scope it can't guess and errors, asking for --scope. With neither --side nor --scope, a single available side is picked automatically; with both server and client available you must pass --side.

point0 trace flags

point0 trace <target> <source> — print the import path from <source> to <target>. This is the command the importer's deny errors tell you to run.

Engine compiler config

The compiler option on the engine (and per side). Set it to false to turn the compiler off for that side; true / an object turns it on. A built engine forces it to false (a built app never compiles sources at runtime).

importer is per-side only, not engine-level. The full engine config surface is on Engine config; the importer options on importer.

The default file filter

By default the compiler claims files matching:

.js .jsx .ts .tsx .cjs .mjs .cts .mts .md .mdx .mdc

It skips ids containing shim: / virtual:, and skips node_modules paths unless the path contains point0. Override with compiler.filter.

Package exports

import { Compiler } from '@point0/compiler' // the class + helpers, via the barrel
import { compilerBunPlugin } from '@point0/compiler/plugin/bun'
import { compilerVitePlugin } from '@point0/compiler/plugin/vite'
import compilerBabelPlugin from '@point0/compiler/plugin/babel'