Talk
Reflections on Context Engineering for MCP Servers
The Model Context Protocol (MCP) has emerged as a standard for connecting AI agents to external tools and data sources. But building an effective MCP server isn't just about exposing APIs—it's about engineering the right context at the right time. When we built the MotherDuck MCP Server to enable agentic data analytics—letting AI assistants like Claude, ChatGPT, or Cursor explore schemas, query databases, and build data applications (`dives`)—we had to realize that context is shaped by a multitude of mechanisms: initial instructions, the tool set design itself, tool descriptions, response structure and length, error feedback, "skill-loading" tools, and sub-agent delegation. Each mechanism involves trade-offs. Eager context loading risks bloat; lazy loading adds tool calls. Rich descriptions help agents self-correct but consume tokens. Getting this balance right means staying conscious of how much context is injected, when, and at what cost (e.g. in # tokens and tool calls to achieve a goal). This talk shares our mental model for context engineering when building MCP servers—shaped by multiple iterations of the MotherDuck MCP Server, but applicable to anyone designing how AI agents interact with tools.
About
Till Döhmen is AI Lead at MotherDuck, where he focuses on building agentic experiences for data analytics. He designed and built the MotherDuck MCP Server, enabling AI agents to query and analyze data through Claude, Cursor, and other MCP clients. Till is also a final-year PhD candidate at the University of Amsterdam, researching AI for data management.