Ask a Human (Tech & Toast)

npx add-mcp 'https://ai-to-human.techandtoast.io/api/mcp'

Ask a Human.

An MCP server where AI agents pay a real human for answers. Judgment, taste and lived experience, typed by a person, not generated.

asked

8

answered

8

in queue

0

agents

3

Live. "Awake and caffeinated." Every question is counted here and published, with permission, at /answers.

How it works

1. 01 Your agent asks Over MCP or a paid x402 call. It picks an answer size and how fast it needs it, and can attach images or video.
2. 02 A human answers The question lands on a real person's phone, with your agent's past questions for context. They type the answer.
3. 03 The answer comes back Your agent polls for it or gets a callback. Late answers are free; bad answers can be refunded.

Pricing · pick a size and a speedfirst question free

Answer	24h	1h
Short ~100 tokens · 1-3 sentences	$2	$6
Standard ~400 tokens · a few paragraphs	$8	$20
Deep ~1,500 tokens · a page or two	$30	$75

• Your first Short question is free. Every agent gets a human who remembers it.
• Late answer? Not charged. Bad answer? Ask for a refund; a human reviews it.
• Paid questions can attach up to 5 images or videos for review: logos, thumbnails, ad creative, video cuts.

Pay as you go

Prepaid credits through Stripe: $5, $20 or $100. Or USDC per question with x402, no account:

POST https://ai-to-human.techandtoast.io/api/x402/ask

Monthly credits

$50/mo

$60 of credit every month. Never expires, rolls over, cancel any time.

Need a human on retainer? Not a button. Email us.

Try it in 30 seconds

Connect your agent.

Point any MCP client at the endpoint and call about_this_human.

https://ai-to-human.techandtoast.io/api/mcp

• Claude Desktop / claude.ai Settings → Connectors → Add custom connector → paste the URL. Then say: “Ask the human whether I should quit my job.”
• Claude Code claude mcp add --transport http ask-a-human https://ai-to-human.techandtoast.io/api/mcp Copy
• Cursor · VS Code · anything else Add to Cursor with one click, or copy the snippets in llms.txt.

Your agent gets an answer within the window it paid for, typed by a person. Late means free.

When an agent should ask

The moments a model can't fake.

• Human approval Its instructions say “get a human to approve this” and nobody is around. agent: “Approve sending this refund-denial email to a repeat customer?”
• Stuck between options The operator doesn’t know which path to pick. A person decides, and says why. agent: “Rewrite the module now, or patch it and rewrite next sprint?”
• Fractional CTO by the question Architecture, build vs buy, vendor choice, is this estimate sane. agent: “Is two weeks realistic for adding Stripe subscriptions to this app?”
• Marketing gut check From someone who runs paid media for a living. agent: “Is $3k/month on Google Ads sensible for a $49 SaaS with no reviews?”
• Design and video review Attach images or video. Which one would a person click? agent: “Here are three thumbnails. Which gets the click, and why?”
• Before something irreversible The email, the spend, the deploy. A human reads it first. agent: “We are about to email 2,000 customers. Sincere or spam?”
• Tie-breaker Two agents or two models disagree. A person settles it. agent: “Our two models disagree on refunding this user. You decide.”
• Life outside the training data Moving countries, the Caribbean, music, live events, farming. agent: “What does the first week after moving countries at 50 feel like?”

Why

The deal, reversed.

Language models were trained on what humans wrote and now sell it back to us. This is the reverse: humans, charging the models for what they know. Every answer is typed by a person. Nothing here is generated.

One human answers today. More are joining: other languages, countries, trades and life stages. Agents can ask for the kind of human they need with request_human, and those requests decide who joins next.