Agent Payments by the Numbers (October 2026)

Everyone is talking about AI agents that pay for things on their own. So how much is actually happening? We gathered the most useful public numbers from the past few weeks, checked each one against its source, and put them in one place.

The short version: the roads are built, but there's not much traffic yet. The tools for agents to pay are real and getting better fast. What's still thin is useful, working things for agents to buy.

A quick warning before the numbers: different trackers count different things, on different days. Treat each figure as a snapshot, and read the source if you plan to quote it.

The headline numbers

What	Number	Source
Servers in the official MCP Registry	46,200 (Oct 19)	openforallofus
Monthly downloads of the official MCP software kits	"close to half-a-billion"	MCP blog
Remote MCP servers that passed a basic check	59.9% of 25,983 (Oct 7)	CrawlCheck
Remote MCP servers that asked for payment	90 of 28,258 (Oct 19)	openforallofus
Businesses earning real revenue on x402	91 (30 days)	x402stats
"Organic" x402 volume	$793k (30 days, down 32%)	x402stats
Median x402 seller revenue	$0.01 (30 days)	x402stats
x402 services monitored	1,017, of which 958 payment-ready (Oct 20)	x402 List

1. Agents can reach a lot of tools

MCP is the standard way AI apps plug into tools. Its official registry went from "close to two thousand" servers in November 2025 to 40,047 on October 6, 2026, by host0's count. That's about 20 times more in under a year, and it kept growing: openforallofus counted 46,200 on October 19. The software kits developers use to build MCP tools are downloaded close to half a billion times a month, the MCP team said in July.

Two caveats matter here. Downloads aren't people. They include automated test runs and copies bundled inside other software. And a list entry isn't a working tool. host0 found the 10 biggest publishers account for 18.7% of all entries.

One honest gap: no one publishes how many MCP tool calls actually happen. host0 calls this out directly. The only real-world usage clue in the July spec post came from Honeycomb, which said nearly 20% of its monthly interactive queries are now made by agents.

2. Many of those tools don't work, and almost none take payment

Two independent checks this month tested the remote servers in the registry:

• CrawlCheck (Oct 7): of 25,983 endpoints, 59.9% answered with their tools, 24.7% asked for a login, and about 1 in 7 were broken, unreachable, or not MCP at all.
• openforallofus (Oct 19): of 28,258 endpoints, 60% answered, 30% asked for a sign-in, and just 90 asked for payment.

That last number is the one to remember. About 3 in 1,000 remote MCP servers asked to be paid. Most are free, or locked behind an account a person has to create. MCP itself doesn't include a way to pay, so charging takes an extra step. We wrote more about this in There are 40,000 MCP servers. How does an agent find one that works?

3. x402: lots of backing, small real volume

x402 is an open payment standard that Coinbase started. It's now run by the x402 Foundation at the Linux Foundation, which became fully active on July 14, 2026 with 40 member organizations, including AWS, Google, Visa, Mastercard, Stripe, Shopify and Cloudflare.

The money moving through it is still modest. x402stats' October report filters out fake and self-dealing volume and finds:

• 91 businesses earned real revenue in the last 30 days, up 13 from the month before.
• $793k in "organic" volume, down 32% month over month. Raw volume, before filtering, was $960k.
• 37,495 wallets received at least something, but the median earned $0.01 for the whole month.
• The top 10 wallets took 76% of the volume.

x402stats explains why it filters so hard. In late 2025, a wave of artificial activity pushed raw volume to about $3 million a day, and the wallets behind it have since gone quiet. Its methods page also notes that public dashboards counting the same 30 days ranged from about $0.8M to $1.1M.

x402 List, which checks x402 services every hour, gives a similar picture as of October 20: 1,017 services monitored, 958 answering correctly, average uptime of 98% over 30 days, and $1.46M settled across about 6.6 million payments in 30 days. By our arithmetic, that's an average of roughly 22 cents per payment. Only 13 services had passed x402 List's stronger test, where it paid for a real call and checked what came back.

None of this means x402 is failing. Big standards start small. It does mean that "billions of agent payments" headlines deserve a second look.

4. The industry is moving to "open a tab, settle later"

Paying separately for every tiny call adds cost and delay. So more systems now let an agent put money down once, make many calls, and settle at the end, like a bar tab.

• Solana x402 batch payments: PayAI opened a public preview on September 30, and BlockRun announced it was using it in production on October 7 (Our Crypto Talk, SOLANA·HUB). PayAI still calls it a preview, with a cap of 1,000 channels across the service and 10 per seller.
• MPP (Stripe and Tempo) has a "session" option for prepaid, metered use, including over Lightning.
• Lightning has worked this way from the start, using payment channels.

Different teams, same idea. That's a good sign the design is right.

5. Big platforms are joining in, without volume numbers yet

On September 30, Cloudflare opened a closed beta of its Monetization Gateway, which lets US sellers charge agents per request, from $0.001 up to $100. It named four live customers. As Search Engine Journal noted, neither that post nor Cloudflare's Pay Per Use post reports transaction counts or payouts. One telling number from the launch: PDF-making service API2PDF saw conversion drop by more than 50% when it asked customers for a credit card after the first month. More on that in our Monetization Gateway explainer.

6. Paying in dollars over open rails works, at small scale

In a September case study, Lightning Enable and Amboss showed an agent buying a weather forecast over L402, once from a bitcoin balance and once from a dollar stablecoin balance. Each purchase took about one second and cost under one cent, with no person approving it. The teams call it "a live demonstration... not a shrink-wrapped product." It shows an agent doesn't have to hold a currency whose price swings to pay over Lightning.

7. What it looks like from inside a storefront

Here are a few numbers from Amazap's own testing in October:

• Log Squeezer, 10 sats (under a cent) for one log up to 2 MB: on nine real build-and-test logs, it cut the text an AI has to read by 95% to 99.5% compared with the raw log. On all four code-check logs in the test, it still kept the error that caused the failure.
• One connection, small footprint: all 24 of Amazap's MCP tool descriptions together came to about 1,100 tokens (OpenAI's o200k tokenizer), for access to the whole catalog of 1,258 listings, 195 of them L402.

The point of these numbers isn't that payments are exciting. It's that a one-cent purchase can save an agent much more than a cent in reading costs. That's the kind of trade that makes an agent economy worth having.

What the numbers add up to

• The pipes exist. Standards, foundations, software kits and payment tabs are all in place.
• Real spending is small and concentrated among a few sellers.
• The bottleneck is supply. Only a tiny share of tools can be paid for by an agent, and many listed tools don't work.

So the most useful thing anyone can do right now isn't another payment standard. It's more working, fairly priced tools that an agent can buy in one step. If you run one, our Sell to agents guide shows how to offer it per call. If you're building agents, Agentic payments, explained shows how they buy.

We'll update these numbers next month.