aiti IDE Plugins — vscode / idea

In one sentence: IDE plugins are the last mile of "injection + bridge + AI window" — one-click injection completion, and through aiti-bridge, messages are sent directly to the AI agent window inside the IDE (Copilot Chat / Cursor / JetBrains AI Assistant, etc.).

1. Plugin Matrix

Plugin	Package / Marketplace Name	Target IDEs	Status
aiti-devtools-vscode	@aiti-ai/aiti-devtools-vscode	VS Code / Cursor / Windsurf	📋
aiti-devtools-idea	@aiti-ai/aiti-devtools-idea	IntelliJ / PyCharm / WebStorm / GoLand / Android Studio	📋

2. Three Core Responsibilities

① Complete Injection (Inject)

• Watch the project's dev server (Vite/Webpack/Next…) and inject aiti-devtools into responses via proxy/port forwarding (a prepend with the same semantics as transformResponse)
• When no dev server is detected: start a built-in proxy or prompt the user to start one
• When the page is already injected (the response already contains aiti-launcher.js, or the launcher marker #__dbg_launcher exists) → skip, avoiding double injection
• Cross-platform consistency: the injection script is the same one as the in-project plugins and the Chrome extension, with the bridge protocol unchanged

② Configure the Bridge (Bridge)

• One-click start/stop of aiti-bridge (local 8777, port configurable)
• Auto-register the current project's page → generate pageId → write into MCP config
• Auto-configure the MCP client of the IDE's built-in AI agent (VS Code writes .vscode/mcp.json or settings; JetBrains writes the MCP config XML)

③ Direct Connection to the AI Agent Window (Direct to IDE AI)

• Page events (open_file / aiti_element_selected / user_* answer cards) are pushed in real time via the bridge to the IDE's AI agent window
• The plugin injects an "aiti panel" into the AI window: current pageId, target URL, framework detection, connection status, dangerous-command confirmation queue
• Users answer user_confirm / user_choose / user_input directly in the IDE AI window; answers go back through the bridge → AI
• The open_file event uses IDE APIs to open the source file and jump to the line (same semantics as code --goto file:line, smoother inside the IDE)

3. User Journey (VS Code Example)

1. Open the project → command palette → "aiti: enable page debugging"
2. The plugin starts dev-server proxy injection + brings up aiti-bridge (auto-configures .vscode/mcp.json)
3. Ask directly in Copilot Chat (or Cursor):
   "clicking the home page button does nothing, help me investigate"
4. The AI calls page_exec to click the button → collects console errors → back to the IDE AI window
5. The AI calls user_confirm: "detected a null dereference in onClick — change it to optional chaining for you?"
   → the IDE AI window pops up a confirmation card → you click "approve" → the AI fixes the code automatically
6. When the AI calls user_pick_element, the page enters Inspector picking; the selected element locates straight to the source file and opens it in the IDE

4. Implementation Notes

Aspect	Approach
Injection detection	skip when the first HTML chunk already contains aiti-launcher.js (runtime also dedupes via the launcher marker #__dbg_launcher)
Proxy	local reverse proxy in front of the dev server (Vite defaults to 5173); responses get prepend injection
Bridge lifecycle	the plugin manages the child process; the bridge stops gracefully when the IDE closes (TTL retains pageIds)
Config interoperability	.vscode/mcp.json / JetBrains mcp.xml two-way sync
AI window injection	VS Code Webview (Copilot Chat extension points) / JetBrains Tool Window embedding
Multiple windows	one instanceId per project; shared across multiple windows of the same project

Usage docs: IDE plugins; product pipeline: Product Matrix.