Your browser, but with an AI layer welded to the chassis

What it does

Gemini Nexus is a Manifest V3 Chrome extension that parks a multi-provider AI assistant inside your browser. It runs as a side panel, an injected floating toolbar for selected text, and—crucially—a local agent loop that can click, type, navigate, and screenshot pages through Chrome DevTools Protocol. It speaks to Gemini Web (reverse-engineered), Gemini API, OpenAI, Anthropic, DeepSeek, OpenRouter, Qwen, Zhipu, and generic OpenAI-compatible endpoints, swapping providers without leaving the extension.

The interesting bit

The project treats the browser as the operating system, not just a chat window. It bundles a full CDP-based control layer (background/control/) so the model can literally drive tabs—hover, fill forms, evaluate scripts, wait for DOM changes—while corralling controlled tabs into a native Chrome tab group. On top of that, it bridges to external MCP servers over SSE, streamable HTTP, or WebSocket, which means stdio-based tools (via a local proxy) can join the same tool loop. That is a lot of plumbing to hide behind a side panel.

Key highlights

• Nine provider backends with per-provider Base URL, API key, and model ID management, including a reverse-engineered Gemini Web driver that reuses your signed-in session.
• Agentic browser control through CDP: navigation, interaction, observation, and script execution with accessibility-tree UIDs for precision.
• External MCP tool integration (SSE/streamable HTTP/WebSocket) with multi-server support and selective tool exposure to keep context windows sane.
• Image and screenshot input including canvas cropping, display-capture from other windows, and floating image detection on pages.
• Context management via summary compression and recent-turn trimming; tab-scoped side panel to reduce noise.

Caveats

• Gemini Web integration is reverse-engineered and documented as fragile; Google can break it without notice and the project includes a manual drift check command.
• MCP stdio servers require an external proxy (the docs recommend MCP SuperAssistant) because Chrome extensions cannot spawn local processes.
• Browser control locks to a target tab group, which is clever but means multi-tab chaos is intentionally constrained.

Verdict

Worth a look if you want an open, self-hosted browser agent that does not phone home to a single vendor and you are comfortable juggling API keys and occasional CDP quirks. Skip it if you need a polished consumer product or if “reverse-engineered web client” sets off your maintenance anxiety.