Nothing is asked of ChatGPT or Gemini: you paste an instructions prompt so the model knows what to write, and the extension watches the reply for tool calls, runs them and pastes the results back. That is also the limitation — the loop depends on the model formatting its call correctly, which is why the instructions prompt is a required step rather than a nicety.
A Chrome and Firefox extension that adds a sidebar to web AI platforms and connects it to your MCP servers through a local proxy. The model writes a tool call in its reply, the extension spots it, runs it against your real MCP server, and inserts the result back into the conversation.
- MCP tools inside ChatGPT, Google Gemini, Perplexity, Grok, Google AI Studio, OpenRouter Chat, DeepSeek, T3 Chat, GitHub Copilot, Mistral AI, Kimi, Qwen Chat and Z Chat
- Tool calls detected automatically in the model's output, executed on click, and their results inserted back into the chat
- Auto-execute and auto-submit modes for running the loop without clicking, and a render mode that formats function calls and results
- Per-tool enable and disable from the sidebar, plus a hover list of available tools and an insert button for the working-instructions prompt the platform needs
- A local proxy that reads a standard `mcpServers` config — including an existing Claude Desktop config file — and speaks SSE, Streamable HTTP or WebSocket, adding CORS and health endpoints for remote servers
The extension from the Chrome Web Store or Firefox Add-ons, plus a local proxy started with `npx -y @srbhptl39/mcp-superassistant-proxy@latest --config ./config.json --outputTransport sse`, or with `streamableHttp` or `ws` instead. The sidebar connects to that proxy — `http://localhost:3006/sse` for SSE, `http://localhost:3006/mcp` for Streamable HTTP, `ws://localhost:3006/message` for WebSocket — or straight to a remote MCP server URL that supports CORS. MIT licensed. Version 0.6.0.
One command — npx -y @srbhptl39/mcp-superassistant-proxy@latest --config ./config.json --outputTransport sse