Mirrow

Mirrow is an AI-powered Electron desktop browser that translates visible website text into Persian using a local LM Studio model. It preserves the original page layout by replacing only text node content.

Features

• Embedded desktop browser with URL entry, back, forward, and reload controls
• Layout-preserving page translation through injected DOM text-node extraction
• Local LM Studio translation through the OpenAI-compatible chat completions API
• Batch translation with progress, partial failure handling, and safe JSON parsing
• Dark premium desktop UI with Translate, History, Settings, and About routes
• Settings and translation history persisted across app restarts
• Safe Electron preload APIs with contextIsolation, nodeIntegration: false, and sandboxing

Tech Stack

• Electron
• React
• TypeScript
• Vite
• Tailwind CSS
• TanStack Router
• TanStack Query
• Electron IPC

Requirements

• Node.js v22.12.0
• pnpm
• LM Studio with a loaded local translation model

Use the project Node version before installing or running:

nvm use v22.12.0

LM Studio Setup

1. Open LM Studio.
2. Load the model: translategemma-4b-it.
3. Start the local server.
4. Confirm the server is available at: http://localhost:1234/v1/chat/completions
5. Run Mirrow.

Default app settings now support both online and local engines. Online mode is compatible with Liara's OpenAI-compatible Chat API.

Online base URL: https://ai.liara.ir/api/6a0ccd2d298429714a4b3e25/v1
Online model: openai/gpt-4.1-mini
Local base URL: http://localhost:1234/v1/chat/completions
Local model: translategemma-4b-it
Temperature: 0.2
Batch size: 20
Target language: Persian

If the online API key field is empty on macOS, Mirrow attempts to reuse Mirook's Keychain API key (com.espitman.Mirook / openai-api-key).

How To Run

nvm use v22.12.0
pnpm install
pnpm dev

For a production build:

nvm use v22.12.0
pnpm build

How Translation Works

When you click Translate Page, Mirrow asks the Electron main process to inject a script into the current BrowserView page. That script walks document.body with a TreeWalker, collects visible text nodes, assigns stable IDs, and stores node references in a page-local map.

The main process sends the collected text to LM Studio in batches. Each model response is parsed as JSON and applied back to the original page with:

node.textContent = translatedText;

Mirrow never rewrites document.body.innerHTML, so page structure, links, images, styles, and interaction handlers are left intact.

Known Limitations

• Some websites may block script injection.
• Dynamic SPAs may need re-translation after route changes.
• Very large pages may take time to translate.
• Local model quality depends on the selected LM Studio model.
• Replacing English text with Persian may slightly affect text wrapping.
• Shadow DOM content is not translated in the first version.

Future Improvements

• Auto-translate after page load
• Translate selected text
• Bilingual hover mode
• Per-site translation preferences
• Translation cache
• OCR for images
• PDF translation
• Multi-language support
• Better RTL handling
• Reader mode
• Extension-like mode
• Custom prompts
• Model selector