Sampling is the reverse relationship of elicitation: instead of asking the user something, the server asks the client to run an LLM completion on its behalf via sampling/createMessage — useful for servers whose tools need a model’s help to finish a task (e.g. summarizing a large document as an intermediate step).

How it works elsewhere

In a normal AI-integrated client, a sampling request would be forwarded to the connected model, which generates a completion that’s sent back to the server to continue its work — entirely automated, with real model calls and real token cost.

How MCPFlo handles it

MCPFlo has no model in the loop by design, so instead of calling an LLM automatically:

1. The server sends a sampling/createMessage request, including the message(s) it wants completed and any generation parameters.
2. MCPFlo surfaces this request to you directly — showing what the server is asking a model to do.
3. You type the response by hand, exactly as if you were the model.
4. MCPFlo sends your typed response back to the server as the sampling result, and the tool call continues.

Why this matters

• Deterministic, token-free testing — you can exercise a server’s sampling-dependent tools without spending real model tokens or getting non-deterministic output, which makes it easy to test specific scenarios (e.g. “what happens if the model says X”).
• Full visibility into what’s being asked — you see the exact prompt/messages the server constructs for the model, which is useful for verifying the server builds sensible sampling requests before it’s ever wired into a real client.
• Testing edge cases — you can deliberately type unusual, malformed, or edge-case responses to see how the server handles them, something that’s hard to force reliably from a real model.

Related

• See Elicitation for the reverse direction — a server asking the user something directly rather than asking for a model completion.
• See Live Notifications for other server-to-client communication that can occur mid-call.
• The bundled test server, @mcpflo/server-everything, includes a tool exercising sampling.