Elicitation
Your agent doesn't need to support elicitation — Meandr handles it.
Some MCP tools stop partway through a call to ask the caller a question. If your agent's client can't answer, the call stalls and fails. Meandr sits in that path: it relays the question to the dashboard, holds the tool call open, and feeds a human's answer back — so an eliciting tool completes even from a client that has never heard of elicitation.
The problem
A tool asks a question your client can't answer.
Elicitation lets a tool pause mid-call to gather one more thing from a human — a target stage, a confirmation, a missing field. It only works if the caller knows how to render the form. Plenty of agent clients don't, and headless service-to-service calls have no one to render it at all.
The question travels backward
Most of MCP runs one way: your agent calls, the server answers. Elicitation is the exception — mid-call, the tool sends a small form schema back toward whoever called it and waits for a reply before it can finish.
Elicitation is optional — many clients skip it
Rendering an arbitrary server-supplied form mid-stream is real work, and plenty of common clients never declare support for it. When one of those is the caller, an eliciting tool has nowhere to send its question.
So the call stalls, then fails
The server waits for an answer its client can't produce, the call hangs until a timeout fires, and the whole thing fails. A tool that depends on elicitation is simply unusable from an agent that can't elicit.
How Meandr relays it
Meandr answers the question for you.
Because Meandr sits inline between your agent and every upstream server, it's already holding both ends of the call when a tool elicits. That position is what lets it step in — capture the form, hold the line, and route the answer back on the same connection.
Relayed to the dashboard
When the client can't take the form — or there's no interactive client at all — Meandr captures exactly what the tool asked for and presents it to a person in the Meandr dashboard, instead of gambling on a client that never declared support.
The upstream line stays open
This is the crucial part: Meandr holds the upstream tool call open the whole time a person is answering. The connection to the server never drops, so the call can still complete once the answer arrives.
A human answers
The form lands in the dashboard, not in one agent's chat, so the right person can answer it — even if they're not the one who happened to trigger the call. They fill it in and it's done.
Fed back on the same call
Meandr splices the answer back onto the held connection and the tool finishes exactly as if the caller had answered directly. From the agent's side, it called a tool, the tool took a little longer, and it got a normal result.
Governed and audited
A server asking your people for input mid-call is exactly what a gateway exists to see. A relayed elicitation runs through the same policy evaluator as everything else, and the captured form and its answer are stored — encrypted — for the record.
Safe by default: a clean decline
Relaying is a deliberate per-server choice. For a server you haven't opted in, Meandr answers an unhandled elicitation with a structured decline sent straight upstream — so the tool resolves immediately instead of jamming the pipeline while it waits for someone who is never coming.
How it works
Tool elicits. Meandr holds. The call completes.
The whole exchange happens on Meandr's side of the connection. Your agent never learns elicitation was involved.
The tool elicits
A form comes back mid-call.
Partway through a call, the upstream tool sends an elicitation request back toward the caller, carrying the form schema it needs answered before it can finish.
You set up: offline elicitation on, per server.
Meandr relays and holds
To the dashboard, line open.
Meandr presents the form in the dashboard and holds the upstream call open within a bounded window while a person answers. The exchange is policy-gated and recorded.
You set up: who can answer, and the policy that gates it.
The answer returns
The call completes normally.
The answer is spliced back onto the held connection and the tool finishes and returns its result. Let the window lapse and the call is released with a clean decline — never left dangling.
You get: a normal tool result, or a clean decline.
Questions you'd ask anyway.
Do I have to turn this on, or does every server's question reach a human?
How long does Meandr hold the call, and what happens when time runs out?
What if my client does support elicitation?
Does this help fully headless, service-to-service calls?
Is a relayed elicitation recorded like the rest of my traffic?
Run eliciting tools from any agent — even one that can't elicit.
Point your agent at one endpoint and you're governed in minutes — no code changes, no proxy to run.
Get startedQuestions? Email us