An Early Bid to Make AI Agents Pay for Web Pages, One Satoshi at a Time

AI agents now read the open web at machine speed, and nobody has agreed who pays for that access. Argentic, a new Show HN project from user Ag0146, is one concrete bid: an HTTP 402 + Bitcoin Lightning toll booth that charges 10 sats (~$0.001) per session for AI scraping traffic. The protocol underneath it is older than the project, and the live question is whether ad-hoc tolls add up to an access layer, or fragment it.

The mechanism predates the pitch. L402 is an open protocol that pairs HTTP's unused 402 Payment Required status with two pieces already common in the Bitcoin Lightning stack: a Lightning invoice (a short-lived request to pay) and a macaroon (a bearer token, like a single-use cookie, that can be narrowed with caveats). On each request, the origin returns 402 plus a Lightning invoice; the client pays; the invoice's preimage (the secret revealed by the network once payment clears) is then used as a session credential. The protocol is stateless, machine-readable, and was published by Lightning Labs on March 11, 2026, as "the internet-native payments protocol for agents."

Argentic is a working instance of that spec. According to its landing page, the project charges 10 sats per session, targets a settlement latency of under 40 milliseconds, and hands back an hour-long token that unlocks unrestricted access. The walkthrough is a Python script: get an invoice, pay it via any Lightning node, poll with the X-Payment-Hash header (the payment's unique Lightning receipt), then attach the returned token as X-Lightning-Token on every subsequent call. It is the kind of integration a single developer can finish in an afternoon, which is part of why the pattern matters.

What matters is the pattern, not the product. Argentic is one toll booth. If it works, the next hundred tolls will look like it: a small server sitting in front of an origin, charging fractions of a cent to AI scrapers, settling on Lightning, and handing out bearer tokens that expire. The Show HN submission, item 49384477, had four points at snapshot, and that is the right calibration: this is an early, narrow signal of a much larger question.

At $0.001 per session, a million requests cost a publisher roughly $1,000. That is a meaningful number for a site that was free last year and a rounding error for a model lab running crawls that burn millions of dollars a day in compute. The price is not yet a market price; it is a placeholder. The interesting variable is not the 10 sats but what happens when someone wants to negotiate volume.

A toll booth is a middleman by design, and that reading is worth holding onto. The toll does not pay the publisher directly; it pays the operator running the toll infrastructure, who is then responsible for forwarding payment to the origin. "No friction, no accounts" is really "someone else decided what counts as access." A bot that cannot pay gets a 402 and stops; a bot that can pay gets a token and goes through. Whether that ends up looking like a healthy market for attention or like an extraction layer depends on who runs the tolls, how transparent their settlement is, and whether publishers can set their own prices or only take the ones they are offered.

Caveats first. Pricing and the under-40-millisecond settlement target are self-reported on the project's own landing page and have not been independently benchmarked in the available references. The landing page does not link a public source repository, a named team, or a relationship to the upstream lightninglabs/L402 reference implementation. The "first machine-native toll booth" framing is positioning, not a verified market claim. None of that changes the mechanism or the pattern, but it does change how much weight to put on any single number.

The question worth watching is the one Argentic did not answer. As agents become routine readers of the open web, who sets the price of an AI-readable page, and where does the money go? The next move that matters is not another toll booth, but a publisher or a platform that decides to either adopt L402-style pricing at scale or refuse to. Until then, the access layer is being assembled in public, one Show HN at a time.