Buzzebees, AI agents
Databricks MCP Server
Lets ChatGPT and Claude ask questions of loyalty analytics, safely and within limits.
How it flows
- 1AI client
- 2OAuth
- 3MCP tools
- 4Databricks / pgvector
- 5Audited answer
The problem
Teams wanted to ask AI assistants about members, campaigns and brands, but analytics data cannot be handed to a chatbot without authentication, limits and an audit trail.
My role
Designed and built the server, the OAuth integration and the tool contracts.
How it works
An MCP server exposes bounded Databricks analytics, a pgvector knowledge base and messaging-provider tools. Clients authenticate through OAuth, every call is logged in PostgreSQL and results are cached in Redis. Both the streamable HTTP and SSE transports are supported, so different AI clients can connect.
Outcome
People ask in plain language and get answers from governed data, without a dashboard request or direct warehouse access.
What I learned
- Bound the tools, not the model. Narrow queries are safer than clever prompts.
- Different AI clients implement OAuth differently. Test each one end to end.
Built with
- MCP
- Python
- Databricks
- PostgreSQL
- pgvector
- Redis
- Auth0
- Azure OpenAI