The first app is a Chrome/Edge Manifest V3 extension. It reuses the user's existing authenticated web sessions, opens model sites as collapsed worker tabs, and mirrors their responses into a single comparison workspace.

Current app

apps/browser-extension

Workspace v1 supports:

ChatGPT
DeepSeek
智谱清言
Qwen
Kimi
Claude (best-effort response adapter)
Gemini (best-effort response adapter)

The default comparison set is ChatGPT, DeepSeek, 智谱清言, Qwen and Kimi.

Workspace architecture

AI Parallel Workspace
        │
        │ launch(prompt, providers)
        ▼
Manifest V3 Service Worker
        │
        ├─ collapsed ChatGPT worker tab ─ Content Script ─┐
        ├─ collapsed DeepSeek worker tab ─ Content Script ├─ response events
        ├─ collapsed Qwen worker tab ─ Content Script ────┤
        └─ collapsed Kimi worker tab ─ Content Script ────┘
                                                         │
                                                         ▼
                                                Workspace panels

Prompts are not transported in destination URLs. Worker pages receive one-time jobs through extension messaging/session storage.

Repository structure

ai-parallel/
├── apps/
│   └── browser-extension/
│       ├── content/
│       ├── workspace/
│       ├── icons/
│       ├── manifest.json
│       └── service-worker.js
├── docs/
│   └── architecture.md
├── .github/workflows/
├── package.json
└── pnpm-workspace.yaml

Shared packages will be extracted only when a second client or genuine cross-app duplication requires them.

Development

npm run check

Load apps/browser-extension as an unpacked extension from chrome://extensions/ or edge://extensions/.

Clicking the extension icon opens or focuses the full-page Workspace.

Security and reliability

Uses only the host permissions required by supported providers.
Reuses existing browser login sessions; it does not store provider credentials.
Does not put prompt text into provider URLs.
Renders mirrored model output as safe text/structured blocks rather than injecting remote HTML.
Provider DOM adapters are best-effort because AI web UIs are not stable public APIs.

See docs/architecture.md for the runtime design and evolution plan.