Return your MCP `tools/list` in deterministic order and set `ttlMs` — it's now a documented prompt-cache-hit lever
Model Context Protocol·medium signal
The new spec adds a `CacheableResult` interface requiring `ttlMs` and `cacheScope` on `tools/list`, `prompts/list`, `resources/list`, `resources/read`, and `resources/templates/list`, and explicitly states servers SHOULD return tools in a deterministic order to enable client-side caching and improve LLM prompt cache hit rates. `cacheScope` is `"public"` or `"private"`, controlling whether shared intermediaries may cache. If your server iterates a hash map to build the tool list, you are silently reshuffling the prompt prefix and busting the cache on every call.