Sampling (MCP)

Sampling is the Model Context Protocol feature that reverses the usual relationship between server and model: instead of only responding to tool calls, an MCP server can ask the connected client to run an LLM completion on its behalf. The name comes from the underlying operation — sampling tokens from a language model.

Concretely, a server handling a tool call might need intelligence mid-task: summarize a 50-page result before returning it, classify an ambiguous input, or decide which of three strategies to take next. Sampling lets it borrow the client's model for exactly that step.

Sampling is a client capability negotiated at connection time. When a server needs a completion, it sends a sampling/createMessage request containing the messages to complete, plus optional model preferences — hints about whether to prioritize speed, cost, or intelligence — a system prompt, and a max token count. The client runs the completion against whatever model it has access to and returns the result.

The specification puts a human in the loop by design: clients should give the user visibility and control over incoming sampling requests, since a server is effectively asking to spend the user's tokens and inject content into a model conversation. The server never learns the client's API key and never chooses the exact model; it only expresses preferences.

Without sampling, any server needing model access would have to ship its own API key and bill its own usage — pushing credential management and inference cost onto every server author and fragmenting model access across the ecosystem. Sampling centralizes both in the client: one key, one bill, one place to enforce policy, while servers stay stateless and credential-free.

This division also makes server logic more portable. The same server gets smarter or cheaper depending on which client connects to it, without a single code change.

Sampling turns servers from passive tool providers into participants that can reason. A research server can iteratively refine queries based on intermediate results; a data-cleaning server can resolve edge cases that defeat regex. Combined with tools, it allows multi-step agent loops to live behind a single MCP tool call.

There is a commercial angle worth noting: a server whose tool calls embed model-quality reasoning is delivering more value per call than a thin API wrapper, which supports a higher per-call price when the server is monetized — on Loomal, anywhere from the $0.01 minimum upward, paid by the calling agent over x402 before the handler runs.

Client support for sampling is uneven — as of mid-2026 many popular MCP clients still do not implement it, so production servers treat it as an enhancement with a fallback path rather than a dependency. Check your target client's documentation before building around it. There is also a latency cost: every sampling round-trip includes a full model completion, so a tool that samples repeatedly will feel slow compared to one that computes directly.