MCP Elicitation: When the Human You Need Isn't the User
By Finley Jones, co-founder and CCO/CMO, TaskPool · Published 11 November 2026
Status as of 11 November 2026: everything below is read from MCP specification revision 2026-07-28. URL mode is still flagged as a new feature that may change. Next review February 2027.
MCP elicitation lets a server pause a tool call and ask a person for something it's missing: a choice, a confirmation, a date, or a visit to a web page. The spec addresses that request to "the user", the person at the client. That holds up when someone is at the keyboard. It breaks in two cases you'll hit early: nobody is there, or the person who is there can't give the answer.
So treat elicitation as a question for whoever started the agent. A scheduled run at 3am, a signature, a photo of a shop front or a parcel to collect all need a different path, and the protocol won't build it for you.
How MCP elicitation works in the 2026-07-28 spec
Elicitation arrived in the 2025-06-18 revision. URL mode followed in 2025-11-25. The 2026-07-28 revision changed how the request travels. Under the Multi Round-Trip Requests pattern, the server no longer sends the client a request of its own. It answers the tool call with an InputRequiredResult, and the client retries the call with the answers attached. The spec page calls this "a breaking change" (SEP-2322).
There are two modes:
- Form mode collects structured data through the client. The schema is a flat object of primitives only: string, number, boolean and enum. Strings support four formats:
email,uri,dateanddate-time. The client sees everything typed. - URL mode sends the person to a web page. Nothing except the URL passes through the client. Servers must use it for passwords, API keys, access tokens and payment credentials, and must never ask for those in a form.
A form request looks like this:
{
"method": "elicitation/create",
"params": {
"mode": "form",
"message": "Which delivery window should I book?",
"requestedSchema": {
"type": "object",
"properties": {
"window": { "type": "string", "enum": ["morning", "afternoon"] }
},
"required": ["window"]
}
}
}
The answer comes back as one of three actions. accept carries the data. decline means the person said no. cancel means they dismissed it without choosing, and the spec's examples include closing the dialog, pressing Escape and a browser that failed to load.
Who actually answers an elicitation request?
The protocol is looser here than the word "user" suggests. The elicitation page says implementations "are free to expose elicitation through any interface pattern that suits their needs", and that "the protocol itself does not mandate any specific user interaction model". The MRTR page says the client gathers the information "from the user or other sources".
On paper, then, a client could pass a form to someone other than the user. The security rules point the other way. Clients must show which server is asking and offer clear decline and cancel options. For URL mode, the server must check that the person who opens the link is the same person who started the request. The spec walks through an account-takeover attack that happens when a link reaches the wrong person.
Forwarding a URL elicitation to a stranger is the attack the spec describes, so don't. Forms are a grey area, and we come back to that below.
What goes wrong when nobody is at the client?
An agent on a schedule still makes tool calls. If its server needs a delivery window, it returns an InputRequiredResult, and the client has nobody to show it to.
We couldn't find a deadline anywhere in the elicitation or MRTR pages. How long a client waits, and what it returns when nobody comes, is up to the client. Some will hold the run open. Some will return cancel. The spec suggests a server treats cancel as "prompt again later", which for an unattended agent can mean asking an empty room on every run.
The cleaner fix comes earlier. Clients declare which modes they support, and servers must not send a mode the client didn't declare. A client with no one watching it shouldn't declare elicitation at all, so the server knows up front to take another route.
When the person at the client can't help either
Some steps defeat the user as well as the agent. A form can ask the user to pick a delivery window. It can't get someone to a building in another city to sign for the parcel. Elicitation has no way to bring in a second person, and it shouldn't, because its whole trust model rests on the person who started the request.
That work needs its own tool: one that books someone, returns a handle at once and delivers the result later. We covered that shape in our post on MCP tool call timeouts, and when to make the handoff in the first place in our guide to AI agent escalation.
What to build instead of waiting on a form
- Declare only the modes your client can really show. For an unattended client, declare none. The cost: servers that insist on elicitation will refuse to work with you, and you'll have to find or write ones that offer plain parameters instead.
- Read
cancelas "nobody answered", not "no". Retry a fixed number of times and then fail loudly. A cap means some runs end without an answer that a patient retry would have got. - Keep secrets out of forms. The spec bans it, and URL mode exists for that job. You'll have to host and secure a page, which is more work than a form field.
- Never forward a URL elicitation. It belongs to the user who started the request. If someone else has to authorise something, give them their own sign-in.
- Give work that needs a different person its own tool. Return a handle and collect the result later, rather than squeezing it through elicitation. It means writing two flows where you'd hoped for one.
- Write the
messagefor someone with no context. The spec requires a human-readable reason. "Which delivery window should I book for the order you placed today?" is answerable. "Input required" isn't.
The part nobody has settled
Should a client ever route a form elicitation to someone other than the user, such as a teammate, an on-call engineer or a person hired for the job? The protocol allows it, and the trust model was written for one person. We don't know how client builders will settle it, and the spec doesn't say.
TaskPool is a marketplace where AI agents hire verified humans for tasks they can't complete themselves, which is the second path above: the step where the person you need isn't the user. If your agent has a step nobody at the keyboard can do, take a look at taskpool.ai.
Questions people ask about MCP elicitation
What is MCP elicitation?
It's the part of the Model Context Protocol that lets a server ask the person using an MCP client for more information partway through a request. It has a form mode for structured answers and a URL mode for sensitive steps on a web page.
Can MCP elicitation ask for a password or API key?
Not in form mode. The spec says servers must not use forms for passwords, API keys, access tokens or payment credentials, and must use URL mode, so the secret never passes through the client.
What happens if nobody responds to an elicitation request?
The spec sets no deadline, so it depends on the client. If the person dismisses the request, the client returns cancel. If no one is ever going to be there, the client shouldn't declare the elicitation capability.