◈OpenTree DashboardSign in

$ date -I # 2026-07-25

Two ways an MCP server reaches back: elicitation and sampling

Almost everything in MCP flows one way: the client asks, the server answers. Two capabilities go the other direction, and they exist because some tool calls genuinely cannot be completed with the information they were handed.

Why a server would need to reach back

Consider publishing with a gate. The model calls the tool with content and no password, because it has no business inventing one. The server can fail, guess, or — with elicitation — ask the human for a passphrase and continue. All three are legitimate designs; only the third neither loses the call nor fabricates a secret.

Elicitation: asking the person

The server pauses mid-call and requests a value, describing what it wants with a schema. The client renders a prompt, the human answers or declines, and the call resumes or aborts. The rules that matter: ask for the smallest thing that unblocks you, make declining a first-class outcome, and never use it for something that could have been a parameter.

Sampling: asking the client's model

The server asks the client to run an inference on its behalf. The appeal is obvious — language understanding without shipping an API key or paying for tokens. The trade-offs are equally obvious: you do not choose the model, the client may refuse, and latency becomes somebody else's problem. Treat a sampling result as untrusted input, because it is.

The draft round-trip pattern

The two compose into a genuinely useful shape: sample to produce a draft, elicit to have the human approve or edit it, then act. For a publishing tool that is "summarise this into a title", "is this title right?", then publish. It keeps the human on the decision without making them do the typing.

What this deployment does instead, and why

Neither is implemented here. Publishing needs no value the caller could not have supplied, and the one place a human decision genuinely matters — content that trips the sensitive-data scan — is handled by refusing with an explicit override rather than by prompting. A refusal the caller must consciously override is auditable; a dialog somebody clicked through at 3am is not.

# blocked, on purpose
POST /sites  →  422 { "error": "sensitive_content", "matches": [ … ] }

# the caller decides, explicitly
POST /sites     { "content": "…", "allow_sensitive": true }

Do all clients support these?

No, and support is uneven. A server that depends on elicitation to function will simply fail on clients without it, so treat both as enhancements with a fallback path.

Is elicitation a security risk?

It is a server-authored prompt rendered in a trusted interface, so it is a phishing surface. Clients mitigate by showing which server is asking; servers should ask for as little as possible, as rarely as possible.

Can I use sampling to avoid paying for a model?

Mechanically yes, and it is worth asking whether that is a fair thing to do to the client's quota. It is best used where the inference genuinely belongs to the user's task.

What should a server do when a user declines?

Return a clean, non-error outcome saying so. Turning a decline into a failure teaches the model to retry, which produces a second dialog.

Keep reading