---
'@modelcontextprotocol/client': minor
---

`Client` now **honours** the server-stamped SEP-2549 `ttlMs`/`cacheScope` cache hints on the cacheable verbs (`listTools()`, `listPrompts()`, `listResources()`, `listResourceTemplates()`, `readResource()`): a still-fresh held entry is served without a round trip. New `CacheableRequestOptions.cacheMode` (`'use'` — the default; `'refresh'` — always fetch and re-store; `'bypass'` — fetch without consulting or writing the cache) gives per-call control. The behaviour is opt-in by hint: a server that sends `ttlMs: 0` (the conservative default this SDK's server stamps) sees byte-identical behaviour — every call fetches.

Entries are automatically scoped by connected-server identity (derived from `serverInfo` after connect, encoded collision-free via `JSON.stringify`); `ClientOptions.cachePartition` is the opaque per-principal slot for `'private'`-scoped entries — set it to your principal identifier (e.g. the auth subject) when one `responseCacheStore` backs several principals. With the default `''` every entry lives at the connected server's shared partition (the safe single-tenant posture). `ClientOptions.defaultCacheTtlMs` (default `0`) supplies the TTL when a result lacks one (e.g. a legacy-era response); the server-supplied `ttlMs` is clamped at 24 h (`MAX_CACHE_TTL_MS`). The list verbs always store the aggregate (so `callTool`'s mirroring/output-validation index keeps working at any TTL); `readResource` stores only when the resolved TTL is positive. `notifications/resources/updated` evicts the cached `resources/read` body for that URI. `ResponseCacheStore` gained `delete(key)`; `InMemoryResponseCacheStore` is now bounded (`{ maxEntries }`, default 512, oldest-first eviction). New exports: `CacheMode`, `CacheableRequestOptions`, `InMemoryResponseCacheStoreOptions`, `MAX_CACHE_TTL_MS`.