Designing MCP tools for agents: Lessons from building Datadog's MCP server

Datadog has redesigned its Model Context Protocol (MCP) server to create an ergonomic observability interface specifically optimized for the unique constraints of AI agents. To prevent context window overload, the team implemented token-efficient data formats, token-based pagination, and SQL-like querying capabilities that allow agents to aggregate data directly. These improvements enhance agent accuracy and reduce operational costs by enabling precise, low-token data retrieval.