Skip to main content
OpenIDEA is a Tauri 2 desktop app. The UI is a React + CodeMirror 6 frontend in the system webview; everything that touches the disk, processes or the network runs in Rust. The Rust side is built so that the same backend runs inside the desktop app and, for remote development, on another machine as openidea-server.

​
The big picture

​
Layers

​
Processes

  • The app process hosts the webview and the whole Rust backend (Tokio runtime plus blocking pools). There is one backend for all windows; each project window owns one Project in the host.
  • Language servers and debug adapters are child processes, started lazily per language and workspace root from languages/<id>.toml. The backend speaks LSP / DAP to them (ide-lsp, ide-dap over ide-jsonrpc). See Code intelligence.
  • Tools are run through their CLIs where that keeps behaviour identical to the command line: git for writes (reads use gitoxide in-process), docker, adb, gh, Maven/Gradle, coding agents over ACP.
  • Plugins run in-process as WebAssembly components, each on its own thread with its own sandboxed store. See Plugin host.
  • openidea-server is the same Host with a JSON-RPC front instead of Tauri. The desktop starts it over SSH, WSL or docker exec and routes a remote window’s commands to it. See Remote development.
  • Local servers for outside clients: the MCP server for AI agents (ide-mcp) and the built-in static web server (ide-web), both on 127.0.0.1.

​
How a request flows

Taking Find in Files as an example:
  1. The search feature allocates a RequestId, creates a Channel, and calls commands.searchStart(request, query, channel) from src/bindings.ts.
  2. Tauri dispatches to search_start in src-tauri/src/commands/search.rs, after the invoke handler has checked project trust.
  3. The command forwards to ide_host::ops::search::start, which registers a CancellationToken for the request and runs the ide-search engine on a blocking thread.
  4. Each file’s matches are sent through the channel as they are found; the command resolves after sending Done.
  5. If the user cancels, the UI calls cancel_request(request) and the token stops the walk.
The same ops::search::start serves a remote window: the frontend’s invoke is routed to remote_invoke, the server runs the identical function, and channel values come back tagged with their channel id.

​
Where to go next

Rust or UI?

Which side a feature belongs on, with examples.

IPC

Commands, errors, streaming, cancellation and events.

Crates

Every backend crate and what it owns.

Frontend

Actions, keymaps, context keys and contributions.