The Deliberate Constraints of MCP Elicitation (lord.technology)

The Model Context Protocol (MCP) has formalized an “elicitation” feature to handle a core problem for AI agents: servers exposing tools (databases, GitHub, cloud APIs) often lack the contextual parameters needed to execute requests. Instead of forcing prewritten configs or risky AI guessing, MCP lets a server pause execution and send an elicitation/create request to the client with a short message plus a JSON schema describing the exact flat, primitive fields it needs. The client renders a simple form, the human supplies the data (or declines/cancels), and the server resumes—creating a collaborative execution model where AI, server, and human negotiate stateful decisions without breaking conversational flow. Crucially, MCP intentionally constrains schemas to flat objects (strings, numbers, booleans, enums, patterns, bounds) and defines three outcomes—accept (with content), decline, and cancel—so intent is precise and UI behavior can be consistent. It also forbids requesting sensitive credentials through elicitation, forcing dedicated secure auth flows. Those design choices trade expressive power (no dynamic, nested, conditional forms) for implementation simplicity and ecosystem adoption: every client can render forms once and support elicitation quickly, reducing misuse and encouraging proper tooling for complex configurations. The result is a pragmatic protocol that prioritizes reliable interoperability and safer human-in-the-loop interactions over maximal flexibility.