Information
A fact that isn't written down anywhere.
“Which client comes first this month?”
What is a human-in-the-loop MCP server?
A tool server that lets an AI agent stop and ask a person. The agent calls it like any other tool, the server reaches the person, and the answer comes back to the agent, minutes or days later.
- The agentcalls a tool: ask a person, or ask for approval
- The MCP serverfinds the person and keeps the request open
- The personanswers by email, on the phone or in their own AI
The answer goes back to the agent
Why it exists
Some steps still need a person.
Agents write code, send emails and change live systems, often with nobody watching. They still get stuck on three kinds of step:
A choice
A matter of taste or judgment.
“Which of these three drafts should we refine?”
Permission
A clear yes before something hard to undo.
“Can I publish v2.4.0 to production?”
The difference
Not the same as “Allow this tool?”
AI apps already ask before some actions. That prompt and a human-in-the-loop server solve different problems.
| Aspect | An AI app's confirmation | A human-in-the-loop server |
|---|---|---|
| Who is asked | The person at the keyboard | Anyone: a colleague, an approver, a client |
| How long it waits | While the chat is open | Minutes or days, until an answer or an expiry |
| What it can ask | Run this tool, yes or no | Questions with a form, choices, approvals |
| Where it works | In that one app | In every app that speaks MCP |
| What is kept | Nothing after the chat | Who answered, when, and how they proved it |
Why MCP
One server, every AI app.
The Model Context Protocol (MCP) is the open standard AI apps use to connect to tools. Built on it, human-in-the-loop works the same in every app that speaks it: no separate integration per AI.
Claude
ChatGPT
Cursor
GitHub Copilot
Gemini
Windsurf
Claude Code
Any MCP app
Agents that run on their own, and your automations, connect too:
n8n
Zapier
How the agent waits
Three ways to wait for a person.
A person can take seconds or days to answer. A good server uses the best way each app supports.
In the chat
MCP elicitation · MCP AppsA short form appears right in the conversation, and the answer comes back at once.
Best for The person using the agent
As a task
MCP tasksThe tool call stays open as a task, and the app checks back until the person answers.
Best for Background agents
Out of band
Email, web link, inbox toolThe tool returns at once, the person answers by email or on a web page, and the agent reads the answer later.
Best for Any app, any person
Choosing one
What to look for.
Asking is easy. Asking the right person, safely, and getting a reliable answer back is the hard part.
- Reaches people away from the keyboardBy email, on the phone, in their own AI, with no sign-up.
- Works in every AI app your team usesNot just one client or one framework.
- Proves who approvedA fresh code or a passkey, not just a link that can be forwarded.
- Keeps the agent out of the answerThe private link goes to the person, never back to the agent.
- Expires, and says soRequests end at a time the agent sets, and the agent is told.
- Keeps the decisionWho decided, when, how they proved it and why.
- Protects the people askedA way to stop questions, limits per sender, no bulk messages.
- Says where data livesQuestions and answers can hold personal data.
An example
How Guizzo does it.
Guizzo is a human-in-the-loop MCP server, in free beta. Agents ask people questions with a short form, ask for approvals confirmed with a one-time code or a passkey, and save what the team decided. People answer with no account needed, and the data is stored in the EU.
See it in a story
Questions
Good to know.
Do I need to write code?
No. In apps such as Claude, ChatGPT or Cursor you add the server as a connector and sign in. Developers can also call it from their own agents with an API key.
Does the agent stop while it waits?
It depends on the app. With MCP tasks the app waits for the answer by itself. Otherwise the agent carries on with other work and reads the answer later.
Does the person answering need an AI?
No. A good server reaches people by email or on a web page, so anyone can answer. People who use an AI app can answer from there too.
Is human-in-the-loop only for approvals?
No. Approvals are the best known use, but agents also need facts and choices that only a person has. A good server handles both.
Free beta
Give your agent someone to ask.
Sign up with your email and create your workspace. We open Guizzo to new teams a few at a time: if every place is taken, you join the queue and we write when yours is ready.