Napi

napi-rs

Write a Node.js native addon in Rust. #[napi] attributes describe the JavaScript surface, napi build compiles one target and emits the loader and the types, and CI fans the binaries out into one npm package per platform.

Read the docs first

Run this before writing any napi config, CI, or publish step. The CLI renamed several commands between v2 and v3, and half the packaging behaviour is not guessable.

npx napi --help          # read the FULL output, never pipe to head or tail
npx napi build --help    # per-command help is much richer than the top level

The full documentation is a separate repo. Fetch it once and grep it locally instead of making dozens of web requests:

bunx opensrc path napi-rs/website
# ► ~/.opensrc/repos/github.com/napi-rs/website/main/content/docs/**/*.en.mdx

The pages that matter most: cli/napi-config, cli/build, cli/artifacts, cli/pre-publish, deep-dive/release, cross-build, more/integrations, more/troubleshooting.

Reference projects:

• https://github.com/napi-rs/package-template — the canonical layout and CI
• https://github.com/Brooooooklyn/Image — a large real addon
• https://github.com/napi-rs/napi-rs — read cli/ source when the docs are thin

Requirements

napi-rs v3 needs Rust 1.88+. The CLI needs Node ^20.17 || ^22.13 || >=23.5; that is a build-time requirement and is unrelated to what the addon supports at runtime.

Scaffold, or set up by hand

npx @napi-rs/cli new my-addon --name @scope/my-addon --no-interactive

napi new copies the maintained Yarn or pnpm template, applies the name and target selection, and can emit a working CI workflow. Prefer it for a new standalone package. For an existing crate or a monorepo, add the four input files below by hand.

Rename later through the CLI so Cargo, the napi config, CI, and the generated binding names stay aligned:

npx napi rename --name @scope/my-addon --binary-name my-addon \
  --repository https://github.com/you/my-addon.git

What the build generates, and what controls it

napi build is a code generator. Never hand-edit its output and never commit it. To change any of it, change the input and rebuild.

Two consequences people get wrong:

napi.targets is a packaging list, not a build command. Each napi build compiles exactly one target, chosen by --target, CARGO_BUILD_TARGET, or the host. Adding a triple to targets creates a package directory and an optional dependency; it does not create a CI job or a binary.

No .d.ts means type-def is off. It ships in napi-derive's default features. If default features are disabled, add it back:

napi-derive = { version = "3", default-features = false, features = ["strict", "type-def"] }

Generated does not mean untracked. The official template commits index.js and index.d.ts so the JavaScript surface shows up in review and consumers of the repo can typecheck without a Rust build. It gitignores *.node and npm/. Committing npm/ leaves stale versions that publish over good ones; the release job recreates it with napi create-npm-dirs.

Committed or not, they are still output. A hand edit is lost on the next build and silently disagrees with the Rust source.

Input files

Cargo.toml

crate-type must include cdylib; Node loads a shared library. Add rlib too when examples/ or benches need to link the crate normally.

[package]
name = "my-addon-native"
version = "0.1.0"
edition = "2021"

[lib]
crate-type = ["cdylib", "rlib"]

[dependencies]
napi = { version = "3", features = ["napi8", "serde-json"] }
napi-derive = "3"

[build-dependencies]
napi-build = "2"

[profile.release]
lto = true
strip = "symbols"

napi-build stays on 2 even for napi-rs v3. The v2-to-v3 migration guide says 3; it is wrong, and the crate has no 3.x. Copy the template.

When the same crate also builds for wasm32-unknown-unknown, gate the napi dependency, because napi does not compile for that target:

[target.'cfg(not(all(target_arch = "wasm32", target_os = "unknown")))'.dependencies]
napi = { version = "3", features = ["napi8", "serde-json"] }
napi-derive = "3"

build.rs

Must sit at the crate root.

fn main() {
    napi_build::setup();
}

package.json

{
  "name": "@scope/my-addon",
  "version": "0.1.0",
  "main": "index.js",
  "types": "index.d.ts",
  "repository": { "type": "git", "url": "https://github.com/you/my-addon" },
  "napi": {
    "binaryName": "my-addon",
    "targets": [
      "aarch64-apple-darwin",
      "x86_64-apple-darwin",
      "x86_64-unknown-linux-gnu",
      "aarch64-unknown-linux-gnu",
      "x86_64-pc-windows-msvc",
      "aarch64-pc-windows-msvc"
    ]
  },
  "publishConfig": { "registry": "https://registry.npmjs.org/", "access": "public" },
  "scripts": {
    "build": "napi build --platform --release",
    "build:debug": "napi build --platform",
    "artifacts": "napi artifacts",
    "version": "napi version",
    "universalize": "napi universalize",
    "prepublishOnly": "napi pre-publish -t npm"
  },
  "devDependencies": { "@napi-rs/cli": "^3" },
  "files": ["index.js", "index.d.ts"]
}

repository must be the real repository. npm provenance validates it, so a leftover template URL fails the release.

v3 config field names, and the v2 names they replaced:

.gitignore

npm/
*.node

index.js and index.d.ts stay tracked, like the template does. Regenerate them with a build; never edit them by hand.

src/lib.rs

#![deny(clippy::all)]

use napi::bindgen_prelude::*;
use napi_derive::napi;

/// `#[napi(object)]` is a value shape. It becomes a TypeScript interface and
/// converts to and from a plain object. Every field must be public.
#[napi(object)]
pub struct EventPayload {
    pub element_id: f64,
    pub event_type: String,
    pub x: Option<f64>,
}

/// `#[napi]` on a struct gives it JavaScript class identity.
#[napi]
pub struct Renderer {
    count: u32,
}

#[napi]
impl Renderer {
    #[napi(constructor)]
    pub fn new() -> Self {
        Self { count: 0 }
    }

    #[napi]
    pub fn set_text(&mut self, id: u32, content: String) -> Result<()> {
        if content.is_empty() {
            return Err(Error::from_reason("content must not be empty"));
        }
        self.count += 1;
        Ok(())
    }

    /// `get_count` becomes the property `count`.
    #[napi(getter)]
    pub fn get_count(&self) -> u32 {
        self.count
    }
}

/// A free function becomes a named export. `Option<T>` becomes `arg?: T | null`.
#[napi]
pub fn parse_json(input: String, pretty: Option<bool>) -> Result<serde_json::Value> {
    serde_json::from_str(&input).map_err(|error| Error::from_reason(error.to_string()))
}

snake_case becomes camelCase, struct names become PascalCase. Override with #[napi(js_name = "…")].

Useful attributes: object, array, transparent, constructor, factory, getter, setter, strict, catch_unwind, namespace, skip_typescript. catch_unwind is not a safety boundary; return Result for expected failures.

Calling JavaScript from another thread

ThreadsafeFunction<Arg, Return> is the only safe way. It queues on the Node event loop.

use napi::threadsafe_function::{ThreadsafeFunction, ThreadsafeFunctionCallMode};

#[napi]
pub struct Renderer {
    on_event: Option<Arc<ThreadsafeFunction<EventPayload, ()>>>,
}

#[napi]
impl Renderer {
    #[napi(constructor)]
    pub fn new(event_callback: Option<ThreadsafeFunction<EventPayload, ()>>) -> Self {
        Self { on_event: event_callback.map(Arc::new) }
    }

    fn emit(&self, payload: EventPayload) {
        if let Some(tsf) = &self.on_event {
            tsf.call(Ok(payload), ThreadsafeFunctionCallMode::NonBlocking);
        }
    }
}

The generated type carries the Node error-first shape:

constructor(eventCallback?: ((err: Error | null, arg: EventPayload) => any) | null)

Prefer NonBlocking. Everything a ThreadsafeFunction stores must be 'static, so convert scoped JS values (Object<'env>, Function<'env, …>) to owned Rust data before crossing the thread boundary, or you get E0521.

Build

npx napi build --platform --release
npx napi build --platform                       # debug
npx napi build --platform --release -t aarch64-apple-darwin
npx napi build --platform -- --locked           # flags after -- go to cargo build

Always ship and benchmark the release build. A debug .node is many times slower and looks like an application bug.

A .node cannot be unloaded. require() calls process.dlopen and Node has no unload, so there is no hot reload for native code. Restart the process after a rebuild; a watcher that only remounts JavaScript keeps the old binary.

In a split workspace, make every path explicit. --package is required when the manifest is a virtual workspace:

npx napi build --platform \
  --cwd packages/addon \
  --manifest-path ../../Cargo.toml \
  --package my-addon-native \
  --package-json-path package.json \
  --output-dir .

Keep the loader and its local .node together. Moving only index.js breaks its relative lookup.

Cross-compiling

Pick exactly one flag; any pair is a hard error. Full matrix in docs/cross-build.

rustup target add <triple> first. macOS and Windows targets are far easier on a native runner; copy what the generated CI does.

CI: one job per target, then one collection job

jobs:
  build:
    strategy:
      fail-fast: false
      matrix:
        settings:
          - host: macos-latest
            target: aarch64-apple-darwin
            build: napi build --platform --release --target aarch64-apple-darwin
          - host: ubuntu-latest
            target: x86_64-unknown-linux-gnu
            build: napi build --platform --release --target x86_64-unknown-linux-gnu --use-napi-cross
          - host: windows-latest
            target: x86_64-pc-windows-msvc
            build: napi build --platform --release --target x86_64-pc-windows-msvc
    runs-on: ${{ matrix.settings.host }}
    steps:
      - uses: actions/checkout@v4
      - uses: actions-rust-lang/setup-rust-toolchain@v1
        with:
          target: ${{ matrix.settings.target }}
          cache: false
          # the action defaults to RUSTFLAGS="-D warnings"; keep lint policy in
          # the crate instead of letting CI deny every warning
          rustflags: ""
      - run: npx ${{ matrix.settings.build }}
      - uses: actions/upload-artifact@v4
        with:
          name: bindings-${{ matrix.settings.target }}
          path: my-addon.*.node
          if-no-files-found: error

Set if-no-files-found: error. Without it a target that silently produced no binary publishes a broken platform package.

Publish

The pipeline, in order. Each command does one thing and none of them covers for another:

napi build (once per matrix row)   ►  upload artifacts
        ▼
napi create-npm-dirs               ►  npm/<target>/package.json, from napi.targets
        ▼
download-artifact                  ►  artifacts/**/my-addon.<target>.node
        ▼
napi artifacts                     ►  copies each binary into npm/<target>/ AND the package root
        ▼
napi pre-publish -t npm            ►  syncs versions, writes optionalDependencies,
                                      publishes every platform package, uploads GH assets
        ▼
npm publish                        ►  the root package

napi pre-publish (alias napi prepublish) never publishes the root package. It runs as prepublishOnly, so the surrounding npm publish does that after it returns. Order is forced: platform packages, then the root, then anything that depends on the root. If the root lands first, an install in that window cannot resolve a binary.

A multi-platform release is not atomic. npm versions are immutable and the CLI cannot roll back. A missing target file is a warning, not an error, so add an explicit CI gate that every configured target directory holds exactly its expected .node before the publish step runs.

To inspect the tarball safely:

npm pack --dry-run --ignore-scripts

Never use npm publish --dry-run as a safety check. npm can still run lifecycle scripts, and a prepublishOnly containing napi pre-publish will publish the real platform packages with real credentials.

Guard local publishing:

"prepublishOnly": "node -e \"if (!process.env.CI) { console.error('CI is the only release path'); process.exit(1) }\" && napi pre-publish -t npm --no-gh-release"

After a partial failure, do not bump the version. Inventory what exists with npm view <pkg>@<version> version, keep the same artifacts, and re-run the same version; already-published packages are skipped.

Never ship every binary in the main package

The most common napi packaging bug. It costs users a multiple-times-larger install and nothing fails, so nobody notices.

napi artifacts writes each collected binary twice, by design. From the official docs:

Native files are also copied to the root package directory so the generated loader can resolve a local binding.

artifacts/bindings-*/my-addon.<target>.node
        │
        ├──► npm/<target>/my-addon.<target>.node   the per-platform package
        └──► ./my-addon.<target>.node              local require, and napi universalize

The root copy is what lets a CI job smoke-test the addon without installing anything:

- run: node -e "require('./packages/native'); console.log('binding OK')"

Then one glob publishes all of them:

"files": ["index.js", "index.d.ts", "*.node"]

node_modules                       254M
├── @scope/addon                   185M   ◄── all six binaries, from the glob
├── @scope/addon-darwin-arm64       23M   ◄── the one that actually loads
└── consumer package               544K

Nothing breaks because the loader tries the local file before the optional dependency, so the root copies silently shadow the per-platform packages.

The rule: never put *.node in files. The official template does not:

"files": ["index.d.ts", "index.js", "browser.js"]

optionalDependencies already delivers the right binary. Keep *.node in .gitignore so a stray local build cannot be published either. Verify with npm pack --dry-run --ignore-scripts: more than one .node in the list means the glob is back.

The loader contract

The generated loader is output, so debug it through its documented interface, not by reading or editing it. In order it tries the explicit override, a local <binaryName>.<platform-arch-abi>.node, the matching optional package, then a configured WASI fallback.

It throws one error whose cause chain holds every native candidate failure. Print the chain instead of reporting "Cannot find native binding":

try {
  require('./index.js')
} catch (error) {
  let current = error
  for (let depth = 0; current; depth++, current = current.cause) {
    console.error(`[cause ${depth}]`, current.stack || current)
  }
}

Then read the real cause:

The target suffix is a selection contract, not a conversion. Never rename a musl binary to -gnu or an x64 binary to arm64.

Turn on both diagnostics when reproducing:

DEBUG='napi:*' RUST_BACKTRACE=full npx napi build --platform --verbose
DEBUG='napi:*' node ./repro.cjs

Consuming an addon: externalize it

A .node is a shared library, not JavaScript. Never let a bundler transform, hash, or inline it. Mark the addon and its platform packages external, then make sure the deployment actually installs them.

external: ['@scope/addon', '@scope/addon-*']         // esbuild
externals: { '@scope/addon': 'commonjs @scope/addon' } // webpack
serverExternalPackages: ['@scope/addon']              // next.config
ssr: { external: ['@scope/addon'] }                   // vite

Externalizing does not package anything. Copy or install the root package and its matching optional dependency into the image, function, or layer, then smoke test one export from the final artifact.

For CommonJS versus ESM, generate the wrapper you need rather than transforming it: --js index.cjs for CJS, --platform --esm --js index.js for real ESM named exports. A CJS wrapper inside a "type": "module" package must use .cjs.

Gotchas

Command names changed in v3. create-npm-dir → create-npm-dirs, napi universal → napi universalize, --cargo-cwd → --manifest-path, --cargo-flags="--locked" → -- --locked. A stale script fails with an unknown-command error, so run npx napi --help after a CLI major bump.

--platform is not optional in practice. Without it the CLI copies a plain index.node and generates no loader.

crate-type must include cdylib, or cargo builds an rlib and the CLI finds nothing to copy.

A stale .d.ts usually means the export is behind a #[cfg(...)], carries #[napi(skip_typescript)], or the dts cache is stale. Delete only target/napi-rs and rebuild.

Write user-facing docs in Rust. Doc comments on #[napi] items become JSDoc in the generated .d.ts, so that is the only place to put them.