MCP adds structured tool output and elicitation
The June 2025 MCP revision lets tools return typed, structured results, lets a server ask the user for missing input, and tightens how servers handle authorization.
The Model Context Protocol published its second specification revision of the year on June 18, 2025.
What changed
- Structured tool output. A tool can now return typed, structured results alongside text, so a client does not have to parse prose to get a number back.
- Elicitation. A server can pause and ask the user for information it needs, such as which account or which date range, instead of guessing.
- Authorization tightened. MCP servers are now classified as OAuth resource servers, and clients must use resource indicators (RFC 8707) so a token issued for one server cannot be replayed against another.
- Batching removed. JSON-RPC batching, added in March, was taken back out.
Why it matters
Structured output and elicitation both push in the same direction: fewer places where the model fills a gap by guessing. A typed result removes one guess. A question back to the user removes another.
There is a third kind of guess that the protocol cannot remove. When a tool returns a column called status with the value A, the agent still has to work out what A means for this business. Typed output tells it the value is a string. It does not tell it the value means "active, under the finance definition, as of last month's close". That meaning has to be written down by the people who own it.