Ask a Human MCP: The Complete Guide to Human-in-the-Loop MCP Servers

Somewhere around mid-2025, a small idea started spreading through the MCP ecosystem: what if, instead of hallucinating an answer, your AI agent could just raise its hand and ask a human?

The idea stuck because everyone who runs agents has lived the failure mode. The coding assistant that confidently invents a database endpoint at 2am. The research agent that fills a gap with a plausible-sounding fact. The automation that fires an irreversible action nobody approved. In every case, one honest question to a human would have prevented the mess – and the agent had no way to ask it.

Today there's a whole family of "ask a human" MCP servers solving this, and they solve it in genuinely different ways. Some route the question back to you, the person running the agent. Some route it to your team via Slack or Discord. And one routes it to a network of vetted domain experts who do the work, not just answer the question. Which one you need depends entirely on who should be answering.

This guide covers all of them – honestly, including the free open-source tools – and gives you a clean way to decide.

The core problem: agents guess when they should ask

Large language models are trained to always produce an answer. That's a feature in a chat window and a liability in an autonomous agent. When an agent hits a gap – missing credentials, ambiguous requirements, a judgment call outside its competence – its default behavior isn't to stop. It's to fill the gap with something plausible and keep going. The errors compound downstream, and by the time you notice, you're debugging false confidence instead of building.

The MCP (Model Context Protocol) fix is structural: give the agent a tool whose entire job is escalation. When the agent is uncertain, it calls the tool, the question routes to a human, execution pauses (or continues on other work), and the answer comes back into the agent's context. The agent doesn't have to be smarter. It just has to know when to ask – and have someone to ask.

Where the tools in this category genuinely differ is the second half of that sentence: who gets asked.

Two kinds of "ask a human": ask yourself vs. ask an expert

Before comparing individual tools, it helps to see the category split, because the tools within each half are more alike than different.

Ask-yourself tools route the agent's question to you or your team – via a markdown file, a CLI prompt, a web UI, a Slack message, a Discord ping. They're lightweight, mostly free and open source, and perfect when you're the person with the answer: your codebase, your credentials, your architectural preferences, your approval on an irreversible action. The agent is borrowing your judgment for a moment.

Ask-an-expert tools route the task to someone who knows more than you do about the thing in question – a researcher, an analyst, a designer, a domain specialist. The agent isn't asking for a quick clarification; it's delegating a unit of work that needs expertise or verification. This half of the category is where Tendem sits.

Most serious agent setups eventually want both. They're not competitors; they're layers.

The ask-yourself tools, compared

ask-human-mcp – the original, file-based

ask-human-mcp by Mason Yarbrough is the tool that gave the category its name. The mechanic is charmingly simple: when the agent is unsure, it calls ask_human(), the question lands in a local markdown file with a PENDING marker, you type the answer, and the agent picks it up and continues. Zero config, cross-platform, works with Cursor and Claude Desktop, full Q&A history preserved as a paper trail.

Best for: solo developers in long coding sessions who want their agent to stop guessing at project-specific details. The author's own framing is apt – it feels like mentoring a sharp intern who actually asks before guessing.

Limits: the human is you, the interface is a text file, and it's synchronous by nature – the agent waits until you answer.

call-a-human-mcp – approvals via Slack, Telegram, or macOS dialogs

call-a-human-mcp extends the same idea to messaging channels and adds a second tool that matters for production agents: request_approval. Before the agent drops a table, deploys to production, or does anything irreversible, it posts an Approve/Deny message to your Slack or Telegram, and it will not proceed until someone clicks a button. Questions work the same way – reply in the thread, the agent continues.

Best for: teams running agents against real infrastructure who want a human gate on destructive actions. The Slack integration means the decision trail lives where the team already works.

ask-user-questions-mcp (AUQ) – structured questions for parallel agents

AUQ is the power-user option: multiple parallel agents can queue single-choice and multiple-choice questions to a separate terminal UI, with question rejection, elaboration requests, native OS notifications, and multi-language support. If you're running long autonomous loops with several agents and want to inject your intent without babysitting any of them, this is the most complete implementation of that workflow.

AskMeMCP and the Discord option – pick your interface

AskMeMCP serves the agent's questions through a self-contained web UI, with tools like ask-one-question, ask-multiple-choice, and challenge-hypothesis. KOBA789's human-in-the-loop does the same job through Discord – the agent posts the question in a channel and mentions you. Functionally these are interface variations on the same pattern: same human (you), different inbox.

MCP elicitation – the protocol-native version

Worth knowing: MCP itself now includes an elicitation mechanism, which lets any server pause execution and request structured input from the current user of the client application. Several of the tools above predate it and some plan to migrate to it. Elicitation is the standards-track future of the ask-yourself pattern – but by definition, it still routes to whoever is sitting at the client. It can't make that person more of an expert than they are.

The ask-an-expert tier: Tendem

All the tools above share an assumption: the person best placed to answer the agent's question is you. Often true. But there's a whole class of agent escalations where it isn't – where the honest answer to the agent's question is "I don't know either."

Is this contact data current? Is this market analysis methodologically sound? Does this deck actually work on a professional design level? Is this scraped dataset complete, or did it silently miss half the source? These aren't clarifying questions. They're work – verification, research, judgment – and routing them to yourself via a markdown file doesn't get them done. It just moves the guessing from the agent to you.

Tendem is the MCP server for that class of escalation. Instead of routing to the current user, it routes the task to a network of 10,000+ vetted domain experts across 30+ skill sets – researchers, analysts, designers, copywriters, engineers – built on Toloka's decade of human-AI collaboration infrastructure. The workflow, from the agent's side, looks like this:

1. You (or your agent, mid-task) describe the work: "Use Tendem to verify these 40 contacts and fill in missing LinkedIn URLs."

2. Tendem scopes and prices the task in chat, usually within about 10 minutes. Nothing is charged before you approve.

3. A vetted expert does the work. The execution is non-blocking – your agent keeps going on other things while the expert works. Most tasks complete within hours.

4. The finished output returns to the same conversation, as structured text plus files, and the agent picks up where it left off.

Two design details matter for anyone comparing this to the ask-yourself tools. First, the async model: file-based and dialog-based tools block until someone answers, which is fine for a 30-second clarification and unworkable for a 4-hour research task. Second, the output is a deliverable, not a reply – verified data, cited research, reviewed designs – which the agent can immediately build on.

In benchmarks across 94 real-world business tasks, this hybrid architecture scored a 74.5% "Good" rating versus 53.2% for human-only freelancers, with 1.8x higher quality than AI-only execution. The gains concentrate exactly where agents fail alone: completeness, factual verification, and the judgment calls that separate plausible from correct.

Side-by-side comparison

How to choose: three questions

1. Do you know the answer? If the agent's question is about your codebase, your preferences, or your approval – ask-yourself tools. If the question requires knowledge or work you don't have on hand, expert tier.

2. Is it a question or a task? "Which auth endpoint do we use?" is a question; a markdown file handles it in ten seconds. "Verify this dataset against primary sources" is a task; it needs hours of someone's time and a deliverable format. Questions go to the ask-yourself layer, tasks to the expert layer.

3. Can the agent afford to wait? Blocking tools stop the agent until a human responds – fine for interactive sessions, a problem for autonomous pipelines. Tendem's async model lets the agent file the task and keep working, which matters once you're running agents that don't have a human watching them.

In practice, a mature agent stack often runs both layers: an approval gate like call-a-human-mcp on destructive actions, and Tendem for the escalations that need expert-grade output. Our broader piece on when to use human experts instead of AI goes deeper on the second layer's decision rules.

Setting up the expert layer

Tendem connects like any other MCP server, no SDK or config files. In Claude: Settings → Connectors → Add Custom Connector, name it Tendem, paste https://mcp.tendem.ai/mcp, authorize via OAuth. Same pattern in ChatGPT, Cursor, Claude Code, Codex, and any client that supports custom connectors. From then on, "Use Tendem to…" in any prompt routes the task to an expert, and the result comes back to the same conversation.

Pricing is per task and shown in chat before anything starts – you approve, trim the scope, or walk away, and nothing is charged until you say yes. New accounts get a $50 starting bonus, and the first three tasks are 50% off, which is enough to run a real escalation through the workflow and see the output quality for yourself. Connect the Tendem MCP here.

Frequently asked questions

What is an ask-a-human MCP server?

An ask-a-human MCP server is a tool that lets an AI agent escalate to a person instead of guessing. When the agent hits uncertainty – missing information, an ambiguous requirement, a risky action – it calls the server, the question or task routes to a human, and the answer returns to the agent's context. Implementations range from local markdown files (ask-human-mcp) to Slack approval buttons (call-a-human-mcp) to vetted expert networks (Tendem).

What's the difference between ask-human-mcp and Tendem?

Who answers, and what comes back. ask-human-mcp routes the agent's question to you through a local markdown file – free, instant to set up, and ideal when you have the answer. Tendem routes the task to a network of 10,000+ vetted domain experts and returns a finished deliverable – verified data, cited research, reviewed work. One borrows your judgment for a moment; the other delivers expertise you don't have on hand. Many agent setups use both.

Does MCP have built-in human-in-the-loop support?

Yes, partially. The MCP specification includes elicitation, which lets a server pause and request structured input from the current user of the client. It covers the ask-yourself pattern at the protocol level. It doesn't cover expert delegation – elicitation can only route to whoever is at the keyboard, so tasks needing outside expertise still require a server like Tendem that connects to an expert network.

Are ask-a-human MCP servers free?

The ask-yourself tools – ask-human-mcp, call-a-human-mcp, AUQ, AskMeMCP – are open source, because the human doing the answering is you. Tendem is priced per task, since a vetted expert is doing real work; every task is priced transparently in chat before anything starts, new accounts get a $50 starting bonus, and the first three tasks are 50% off.

Which AI tools support ask-a-human MCP servers?

Any MCP-compatible client. The open-source tools are most commonly used with Cursor, Claude Code, and Claude Desktop, since their audience is developers in coding sessions. Tendem connects to Claude (web, desktop, Claude Code, Claude Cowork), ChatGPT, Cursor, Codex, and any client that supports custom connectors – setup is a one-minute OAuth flow with no config files.

Related Resources

• Human-in-the-Loop: The Complete Guide – the architecture behind AI + human systems

• How to Delegate Work to Human Experts from Claude or ChatGPT – prompt patterns for the expert layer

• Tendem Now Works Inside Claude, ChatGPT, and Your Other AI Tools – the launch announcement

• The True Cost of AI Hallucinations in Business Data – why agents guessing is expensive

• Tendem MCP – connect the human expert layer to your agent