Language
Swiftraycast.com
Writes the macOS host app that handles windows, hotkeys and the menu bar.
TypeScriptraycast.com
Is the language of the front end, the Node core and third-party extensions.
UI library
AppKitraycast.com
Provides the native macOS layer of the app.
Why: They weren't confident SwiftUI wouldn't raise crash rates or hurt performance.
Reactraycast.com
Builds the shared UI for every window and the extensions API.
Why: One web UI codebase serves both platforms while the native shells keep full access to OS APIs.
WebKitraycast.com
Renders the UI on macOS through the system WKWebView, with WebView2 doing the same on Windows.
Why: They didn't want to bundle Chromium on macOS when the system's WebKit was already there.
Full-stack framework
.NETraycast.com
Runs the Windows host app, written in C# with .NET 8 and WPF.
Build and runtime
esbuildraycast.com
Transpiles extension TypeScript inside the ray CLI.
Is the language of the ray CLI that extension developers use to build and publish.
Data fetching
JSON-RPCraycast.com
Carries messages between the extension runtime and the app over standard streams.
Why: It is lightweight, well supported, fast enough and easy to debug because it is plain text.
Backend framework
Node.jsraycast.com
Runs one long-lived process with database access, the extension runtime and other services.
Why: Node is the shared layer for both platforms, so feature work is done only once.
Rustraycast.com
Powers the shared data layer, cloud sync and the file indexer.
Why: They use it where performance or portability matter more than convenience.
AI model
OpenAIraycast.com
Was the first and for a while only model provider behind Raycast AI.
Anthropicraycast.com
Supplies the Claude models offered in Raycast AI.
Why: No single model is best at speed, intelligence and context window at once, so users pick per task.
The original Mac app was written in Swift with AppKit. The team skipped SwiftUI because they weren't confident it wouldn't raise crash rates or hurt performance, two things a launcher you open dozens of times a day can't afford.
Sources:raycast.com
Raycast wanted extensions to feel first-class, not like mini-apps in a web frame. Developers write React and TypeScript, and a custom React reconciler turns their components into a JSON render tree that the Swift side draws with AppKit, diffing each render with JSON Patch.
The first attempt ran extensions on JavaScriptCore, but developers couldn't use their favorite npm packages. So Raycast moved to Node.js, running each extension in its own worker thread with memory limits and crash isolation, and talking to the app over JSON-RPC 2.0.
Sources:raycast.com
To ship on Windows, Raycast rewrote the app from scratch. Each platform keeps a native host (Swift and AppKit on macOS, C# with .NET 8 and WPF on Windows) that loads one shared React and TypeScript front end in the system web view and supervises a long-lived Node.js process holding the business logic.
They looked at Electron, Tauri, Flutter, Qt and React Native for desktop. Electron made deep OS integration painful and meant bundling Chromium on macOS, and Tauri gave less native control and felt too young to bet the company on. The trade-off is that Raycast now maintains its own IPC, and memory use went up compared with v1.
Sources:raycast.com
Rust is reserved for the places where performance or portability matter more than convenience: a data layer that the iOS app can share, cloud sync and a file indexer running as its own process.
The indexer exists because Windows has no system-wide index good enough for Raycast. On Windows it reads the NTFS Master File Table directly to scan drives quickly.
Sources:raycast.com raycast.com
Raycast AI launched on OpenAI models only. In April 2024, after the community kept asking for more models, Raycast added Anthropic, Perplexity and open models served by Groq, arguing that no single model wins on speed, intelligence and context window at once.
Sources:raycast.com
Found on the homepage with the open Wappalyzer rules (HTML, scripts and headers). It shows what runs in the browser, not the whole backend.
Checked on 2026-10-29. Every tool links to the public source that shows it.
