MCP-powered research for AI agents
An MCP server that brings Lux Research’s proprietary research into the AI agents clients already use: Copilot, ChatGPT, and Claude.
My role: Product + EngineeringThe problem
Clients were already working in AI tools like Copilot, ChatGPT, and Claude. We wanted to meet them where they are, so they didn’t have to switch to a separate portal, and to layer more recent news over Lux analysts’ research.
My role
I led a team of two. I also decided what it should do: the feature roadmap came out of conversations with Lux analysts and clients, and both use it.
What I built
An authenticated MCP server that gives AI agents access to Lux’s proprietary research.
- A search tool over Lux research, using hybrid search with embeddings
- Agents bring in recent public news themselves; the server provides the proprietary research to layer it on
- A different way in for each platform: a declarative agent for Microsoft Copilot, a plugin for ChatGPT, and a connector for Claude
- In progress: a
create_presentationtool that builds branded presentations from a slide template deck
Built with Python and FastMCP, with OAuth through Auth0. Hybrid search runs on a vector database with BAAI/bge-base-en embeddings, and the server runs on AWS ECS.
Outcome
It’s live with clients. Building it taught me a ton about authenticated MCP servers, and about the quirks of connecting to each of the big model providers and Microsoft.