Comprehensive best-practice guide for building, securing, testing, and shipping Tauri v2 desktop/mobile apps (Rust core + any web frontend — React, Next.js, Vue, Svelte, vanilla). Covers IPC (commands/events/channels), the v2 capabilities & permissions security model, CSP, state management (Manager/

Install

openclaw skills install @anjasta-tarigan/tauri2-best-practice

Tauri v2 Best Practice

Tauri v2 is a Rust-core, webview-frontend framework for desktop (Windows/macOS/Linux) and mobile (iOS/Android) apps. This skill is deliberately overkill: it encodes the parts of Tauri v2 that are easy to get wrong, silently insecure, or scattered across many doc pages, so you don't have to relearn them from scratch every session.

Every code example and claim in this skill and its reference files is grounded in the official Tauri v2 docs (v2.tauri.app), docs.rs/tauri, or the tauri-apps/tauri-docs GitHub source. Reference files carry per-topic source URLs at the top — verify against them (via web search / fetch) before shipping anything security-critical, since Tauri v2 is still evolving and APIs move between minor versions.

0. First, orient yourself

Before writing anything, check:

  1. Version. src-tauri/Cargo.tomltauri = "2.x" and tauri.conf.json's $schema pointing at a v2/2.0.0 schema URL confirms v2. If you see tauri.allowlist in tauri.conf.json, that's v1 — stop and flag it; the allowlist system was fully replaced by capabilities in v2, and the two are not interchangeable.
  2. Project shape. src-tauri/ (Rust core: Cargo.toml, tauri.conf.json, src/, capabilities/, optionally binaries/ for sidecars) + a frontend root (any framework, or none — vanilla HTML/JS is fully supported).
  3. What the user actually needs. Route to the relevant reference file below rather than dumping everything — but read security-capabilities.md proactively any time the task adds a new command, plugin, or filesystem/shell/network access, even if the user didn't ask about security.

1. Reference map — read before you write code

If the task involves...Read
#[tauri::command], invoke(), emit/listen, streaming data, Channelreferences/ipc-commands.md
capabilities/*.json, permissions, scopes, CSP, withGlobalTauri, isolation pattern, "why is invoke undefined/failing silently"references/security-capabilities.md
app.manage(), tauri::State, sharing data across commands/threads, async mutex vs sync mutexreferences/state-management.md
Writing a custom plugin, embedding an external binary (Python/Node/Go/Rust CLI) as a sidecar, tauri-plugin-shellreferences/plugins-sidecar.md
Multiple windows, WebviewWindowBuilder, window labels, mobile multi-window, window/webview split, system tray, menusreferences/windowing-multiwindow.md
tauri build, code signing (macOS notarization / Windows Azure Key Vault), the updater plugin, latest.json, GitHub Actions release pipelinereferences/bundling-updater-cicd.md
Unit tests, tauri::test::mock_builder, @tauri-apps/api/mocks / mockIPC, WebDriver / tauri-driver / WebdriverIO, CI test matricesreferences/testing.md

Each reference file is self-contained with runnable code, gotchas, and a "sources" section. Files run 200–500 lines — read the whole file for the topic you're touching rather than skimming, since the gotchas are usually in the second half.

2. The eight rules that cause 90% of Tauri v2 bugs

These are condensed here because they're cross-cutting and worth having in working memory even before you open a reference file.

  1. Nothing is allowed by default. Tauri v2 is default-deny. A window with no capabilities/*.json entry for a command cannot call it — the call fails at runtime (often silently, or as a vague "<command> not allowed" IPC error), not at compile time. If a frontend invoke() call "does nothing," check capabilities before touching the Rust code.
  2. Capabilities are matched by windows label. A permission granted in a capability file only applies to windows whose label matches the windows array (glob-capable). Adding a new window (e.g. a settings window) without adding its label to the relevant capability is the #1 cause of "works in main window, broken in the new one."
  3. Plugin permissions are separate from core permissions, and are namespaced plugin-name:permission-id (e.g. fs:allow-read-file, shell:allow-execute). Installing a plugin's crate + npm package is not enough — you must also grant its permissions in a capability file, and this has no compile-time check.
  4. State is auto-wrapped, don't double-wrap. app.manage(x) makes x retrievable as State<'_, X> from any command. Tauri does not require you to wrap it in Arc yourself for the State extractor to work across threads — do that only if you're independently sharing it outside Tauri's state system. Do wrap the inner data in Mutex/RwLock (or an async tokio::sync::Mutex if you need to hold the lock across .await) since managed state must be Send + Sync.
  5. window.__TAURI__ is opt-in. app.withGlobalTauri defaults to false in v2. Any code (including a plain bundled index.html with no bundler) that references window.__TAURI__.core.invoke instead of importing from @tauri-apps/api will silently fail unless you've explicitly set "app": {"withGlobalTauri": true} in tauri.conf.json.
  6. Sidecar filenames must carry the target triple, e.g. my-sidecar-x86_64-pc-windows-msvc.exe, matching rustc -Vv | grep host (or rustc --print host-tuple on Rust ≥1.84). Command.sidecar() on the JS side takes the logical name from bundle.externalBin, not the suffixed filename.
  7. Creating a window synchronously inside a command can deadlock on Windows (a known WebView2 issue). Always create windows from an async fn command, or spawn a thread/async task, when the trigger is IPC.
  8. CSP defaults matter. tauri.conf.json's app.security.csp is enforced; setting it to null disables CSP entirely (don't, except transiently while debugging). A restrictive CSP plus a locked-down capabilities set is what actually makes "no Node.js/Electron-style full OS access from the webview" true in practice — don't rely on capabilities alone if the CSP is wide open.

3. Working style for this skill

  • When reviewing existing Tauri code, actively look for capability/permission mismatches (rule 1–3) and sidecar target-triple issues (rule 6) even if not explicitly asked — these are the most common silent-failure classes.
  • When scaffolding new commands/plugins/windows, always show the matching capabilities/*.json diff alongside the Rust/JS code — don't leave permission-wiring as an exercise for the user.
  • Default to the principle of least privilege: prefer narrow, per-command permissions and per-window capability files over *:default blanket grants, and say so explicitly if you do reach for a blanket grant (e.g. rapid prototyping).
  • Cross-check anything version-sensitive (plugin APIs churn between Tauri 2.0/2.1/2.2+) against current docs via web search rather than asserting from memory, and say so if something looks like it may have shifted since the reference file was written.