What is MCP elicitation?
MCP elicitation is a feature of the Model Context Protocol (MCP). It lets a server pause in the middle of a task and ask the user for more information. When a required detail is missing, the server does not simply fail. Instead, it sends a structured request back through the client, the user answers, and the task moves forward. Added to the MCP specification in mid-2025, the feature turns a one-shot tool call into a short, controlled exchange.
The request is server-initiated but user-facing. The server sends an elicitation/create message. It carries a plain-language prompt and a small schema that lists exactly which fields the server needs. The client turns that into a form or dialog, collects the answer, and returns it. The user stays in control. They can provide the data, decline, or cancel the whole request.
In plain words
Think of an online checkout that suddenly needs your shipping address. Instead of throwing an error and making you start over, it pops up a small form. You type the address, and the order goes through. MCP elicitation is that pop-up for AI agents. When the agent hits a gap, it asks you one focused question instead of guessing or giving up.
How it works
Elicitation follows a simple, predictable loop:
- The server asks. While handling a request, it sends
elicitation/create. The message says what it needs and why, plus arequestedSchema: a short, flat list of fields such as a name, a number, or a pick from a few options. - The client presents it. The client renders the request as a UI element and shows which server is asking. It never answers on the user's behalf.
- The user responds. The user can accept and submit the requested data, decline the request, or cancel it entirely. All three are valid outcomes the server has to handle.
- Work continues. On accept, the structured answer flows back to the server, which resumes the task with the missing piece in hand.
The schema is deliberately kept simple, primitive fields only, so the client can render a clean form and the user can see exactly what they are about to share.
Why it matters
For anyone building agents on top of MCP, elicitation removes a common failure mode and keeps a person in control at the right moment.
- Fewer dead ends. An agent no longer has to abort or hallucinate a value when one input is missing. It asks, gets the answer, and finishes the job.
- Human in the loop by design. Irreversible or high-impact actions, deleting data, spending money, deploying, can pause for an explicit confirmation with context instead of running blind.
- Better agent UX. One targeted question at the right time beats a long upfront form that asks for things the task may never need.
- A cleaner alternative to guessing. Structured input with a schema is safer and easier to validate than parsing a free-text answer buried in the conversation.
What to watch out for
- Never request secrets. The spec is explicit: do not collect passwords, API keys, or other credentials this way. Use a proper auth flow instead.
- Keep the trust boundary clear. A malicious server could word a request to trick users into oversharing. Clients should show who is asking and make declining easy.
- Keep schemas small. Elicitation is for one missing detail or a confirmation, not a full config wizard. Deeply nested schemas are poorly supported.
- Handle decline and cancel. Users will say no. Treat "declined" and "cancelled" as normal paths, not errors.
- Do not confuse it with related features. Elicitation gathers user input; sampling asks the model to generate; tool permissions authorize a call.
Related articles:
- MCP under the microscope: how AI agents talk to tools - How the Model Context Protocol connects agents to tools, and the risks that come with it.
- What's an agent? - The short version: an AI that handles a task on its own once you give it the goal.
- What is agentic AI? - Why human-in-the-loop checkpoints matter once AI starts acting on its own.
- What is prompt injection? - The trust-boundary problem behind every server-to-user request.
Want to stay one step ahead?
Don't miss our best insights. No spam, just practical analyses, invitations to exclusive events, and podcast summaries delivered straight to your inbox.