MCP Caching
MCP 2026-07-28 makes caching a first-class performance feature for list and resource-read results. Responses can include ttlMs and cacheScope.
ttlMs
ttlMs tells a client how long the result should be considered fresh. A zero or non-positive value means the result should not be served as fresh from cache.
cacheScope
private means a result should stay scoped to the requesting context; public can allow broader sharing when the data is truly identical and safe across users.
Why tool-list caching matters
Repeatedly listing a large server's tools on every agent turn adds latency and can destabilize upstream prompt caching. A stable cached catalog avoids unnecessary round trips.
Be conservative with user-specific data
If a tool or resource catalog changes based on permissions, tenant, role, or user identity, use private caching and a short TTL unless you have a robust cache key that captures the authorization context.