Dictionary term
MCP Sampling
A deprecated MCP client feature that lets a server request a completion from the client's LLM, so the server can use model reasoning without holding its own API key. It was deprecated in spec revision 2026-07-28 and remains in the spec only for a deprecation window.
Deprecated
Deprecated in MCP spec revision 2026-07-28 (SEP-2577), together with Roots and Logging. It still works during a deprecation window of at least twelve months, but new implementations should not adopt it; the spec's suggested migration is to integrate directly with LLM provider APIs.
In plain English
An older MCP feature that let a tool borrow the AI app's own model to do some thinking. The MCP standard now marks it as on its way out, and new tools should call an AI provider directly.
Analogy
Like a contractor borrowing the homeowner's tools instead of buying their own: the server asks the client to do the thinking.
Why it matters
You will still meet sampling in existing MCP servers and clients, but new servers should not rely on it: the spec recommends calling an LLM provider API directly instead, and clients may drop support once the deprecation window ends.
In practice
An older summarization server asks the client's model to condense a fetched document; a server written today would call a model provider's API itself instead.
Related terms
Sources
Browse all terms →Last verified