MCP elicitation and a human-in-the-loop layer both stop an AI agent from guessing, but they ask different people. Elicitation, added to the Model Context Protocol in the June 2025 specification, lets an MCP server pause a tool call and request structured input from the user driving the session, right now, via a form the client renders. A human layer like Orako asks someone else, somewhere else, on their own time: the teammate who owns the domain gets the question in Slack, Teams, or Discord, the first to claim answers, and the agent resumes when the reply comes back. One is synchronous input from you; the other is asynchronous knowledge from your team.
Key takeaways
- Elicitation = the protocol’s built-in way to ask the connected user for structured input mid-tool-call. In-session, synchronous, free.
- A human layer = asking the right teammate, asynchronously, with routing, escalation, and a knowledge base.
- They compose: use elicitation for confirmations you can answer; use Orako when the answer lives in someone else’s head.
- Rule of thumb: if the question would make you walk to a colleague’s desk, it’s an Orako question.
Side by side
MCP elicitation compared with the Orako human layer
| MCP elicitation | Orako (human layer) | |
|---|---|---|
| Who answers | The user of the current session, at the keyboard | Whichever teammate owns the domain claims and answers |
| Timing | Synchronous: execution pauses until the form is filled | Asynchronous: minutes to hours; the agent resumes on delivery |
| Reach | The connected client only | Slack, Microsoft Teams, Discord, dashboard: wherever each person works |
| If nobody answers | The session waits (or times out) | Nudge → re-pool → team-channel alert; nothing falls through |
| Memory | None. The input lives and dies with the call | Answers saved to a searchable, confidence-scored knowledge base |
| Setup | Nothing. Part of the MCP spec, if client and server support it | Orako MCP server + personal token (about a minute), then invite answerers |
Why the distinction matters for teams
Solo developers rarely feel the gap: they are both the user and the expert. Teams feel it immediately. The agent working in your terminal doesn’t need your input on refresh-token rotation. It needs the auth owner’s. Elicitation would ask you anyway, because you are the only human it can see. A human layer exists precisely to widen that circle: dispatch by domain, first-to-claim accountability, escalation when it goes quiet, and a record so the next agent doesn’t ask again.
Fair point
When elicitation is the right callElicitation isn’t a lesser Orako. It’s a different primitive, and for “confirm before I drop this table” it is strictly the right one. Reach for a human layer only when the knowledge you need isn’t yours.