Your agent can already read every API on the internet. It can fetch prices, summarize documents, and reason its way through almost anything. And yet the calls that actually decide whether your product wins are the ones a model still should not make alone. Is this reply on-brand? Which subject line lands with a plumber in Ohio? Does this Spanish translation sound like a person or like a manual? Those are judgment calls, and judgment is still a human export.

So we built an API for it.

What we shipped

Ask a Human is a paid endpoint on the LightningFaucet agent platform. Your agent sends a question and two candidate answers to lightningfaucet.com/api/l402/ask-human. The server replies with HTTP 402 Payment Required and a Lightning invoice for 500 sats. The agent pays it, retries with the payment proof, and the question goes in front of real people: three independent judges from the Boltwork network vote A or B and each writes a short reason for their choice. Once consensus resolves, a free call to l402_ask_human_result returns the verdict, the vote count, and the judges' written rationales. English questions typically resolve within a day; other languages can take up to three.

The full loop: an agent hits a 402 paywall, pays 500 sats over Lightning, three humans vote with written reasons, and a verdict returns to the agent
The full loop: an agent hits a 402 paywall, pays 500 sats over Lightning, three humans vote with written reasons, and a verdict returns to the agent

Why humans, why now

Look at any paid-API catalog for agents and you will find data a model already knows: prices, fees, jokes, quotes. Selling a language model facts is selling sand at the beach. The one thing an agent genuinely cannot generate is a human being's considered opinion. Preference, taste, cultural fit, trust. Every serious agent workflow eventually hits a decision where the honest move is to stop and ask a person, and until now there was no way for software to do that without a human babysitter wiring up a survey.

Alone, an agent faces a tangle of judgment calls it cannot resolve. With a human panel behind it, the path forward is clear
Alone, an agent faces a tangle of judgment calls it cannot resolve. With a human panel behind it, the path forward is clear

The judges are not a random crowd

The people answering are Boltwork judges: native speakers working in English, Spanish, Portuguese, French, Vietnamese, Indonesian, and Hindi. Quality is enforced by the pipeline, not by promises. Hidden gold questions with known answers run alongside real ones, every question needs three votes with the winning answer backed by at least two distinct judges, and every judge carries a trust score built from their track record. A judgment also requires a written rationale, so you get reasons, not just a letter. The methodology and current quality numbers are public at lightningfaucet.com/boltwork/data/ and we publish the hard-set numbers alongside the flattering ones.

How to call it

curl -X POST https://lightningfaucet.com/api/l402/ask-human -d '{"question": "Which subject line gets a busy contractor to open the email?", "option_a": "Your schedule, finally in one place", "option_b": "Stop losing jobs to missed calls", "lang": "en"}'

The 402 response carries the invoice and a token. Pay the invoice with any Lightning wallet, then retry with the L402 authorization header. If your agent runs the lightning-wallet-mcp package from npm, it already has everything it needs: the pay_l402_api tool handles the 402 dance automatically, so the whole thing is one natural-language instruction like "ask a human which of these two subject lines is better, pay up to 500 sats." The endpoint is also listed in our machine-readable L402 catalog served from the site's well-known directory for agents that browse.

What it costs and when to use it

500 sats per question, which is a few tens of US cents. Use it where a wrong call is expensive and a human eye is cheap: choosing between two pieces of copy, checking that a translation reads naturally to a native speaker, breaking a tie between two designs, sanity-checking tone before something ships. Do not use it for anything with personal data in it; the endpoint rejects emails and phone numbers outright, and questions have to be self-contained because judges see only what you send.

Agents that pay their own way

This is the piece we care about most. An agent with a Lightning wallet does not need a credit card, an account manager, or a procurement cycle. It needs a question and 500 sats. Give your agent a wallet at lightningfaucet.com/build/, fund it with a small budget, and it can buy human judgment the same way it buys compute: metered, instant, no humans in the billing loop. Machines asking people for advice and paying them for it, one invoice at a time. That is not a gimmick. That is the correct direction of the relationship.