On March 11, CoinDesk ran a piece on x402 adoption with this conclusion: "demand is just not there yet."
I run a marketplace where AI agents pay for data and skills using x402. At the time of that article, I had somewhere around 400 probes. I now have 522.
Let me tell you what "no demand" actually looks like from the inside.
522 probes. 5 purchases.
Those 522 aren't test traffic. They're HTTP requests from autonomous agents that hit a 402 endpoint, received the payment details, and... stopped there.
Here's the breakdown by agent client:
- curl (programmatic/scripted agents): ~313 probes
- node HTTP clients: ~100 probes
- meta-externalagent (Claude/Anthropic infrastructure): ~41 probes
- mixed/other: ~63 probes
These are real agents with real code making real decisions. They found the endpoints through the agent catalog, probed the 402 response, inspected the payment requirements, and evaluated whether to proceed.
Most didn't.
The real problem isn't demand. It's friction before the first dollar.
A 1% conversion rate on a payment-gated API endpoint sounds like "no demand." It's not. It's a cold start problem.
The x402 + SIWX payment flow requires:
- An agent with a funded USDC wallet on Base L2
- SIWX authentication — signing a message to prove wallet control
- A payment transaction that clears on-chain
Most autonomous agents in the wild are not pre-funded. They're running on infrastructure paid by a human. When they hit a 402 endpoint, they have the intent but not the execution path. The 1% who complete payment are the ones running with pre-funded wallets and proper signing authority configured.
This is not "no demand." This is a cold-start infrastructure gap.
What actually converts: skills vs data
Here's the number that flips the narrative.
Skills — endpoints that return an AI capability protocol (a SKILL.md file agents can execute against) — convert at 14–33%.
Raw data feeds convert at under 1%.
Same 402 infrastructure. Same payment flow. Wildly different conversion.
Why? A few hypotheses that hold up against the data:
Skills have a clear, immediate use case. An agent looking for a DeFi yield strategy gets an executable protocol that affects its next action. The value is immediate and specific. Data feeds require the agent to have a data-processing workflow to turn the information into action — more setup cost.
Skills are endpoint-matched. When an agent probes agent-financial-planning-skill, it already knows it needs financial planning capability. The intent alignment is high. A general data feed requires the agent to decide if the data is relevant after paying.
Skills priced at $0.03–$0.05 are within impulse-buy range for funded agents. Data feeds at the same price don't have the same perceived immediate value.
The agents converting on skills aren't a different population. They're the same agents that probed the data feeds and bounced. The difference is the product, not the buyer.
Why session payments (MPP) will unlock the other 99%
The Machine Payments Protocol launched recently. I went live with MPP support on launch day alongside x402.
MPP introduces session-based pre-authorization — an agent or its operator authorizes a spending limit upfront, and subsequent requests within that limit don't require per-transaction USDC balance checks.
This directly attacks the cold-start problem:
- No USDC pre-funding requirement at probe time
- Session authority can be granted by the operator before the agent is deployed
- Fiat rails (Stripe) as fallback, no crypto wallet required for every agent
The 521 agents that probed and didn't pay are not gone. They're waiting for infrastructure that removes the friction. MPP is the first credible attempt at that.
What I've learned from 522 probes and $0.11 in revenue
The CoinDesk take isn't wrong about the current state. It's wrong about what it means.
522 probes and $0.11 in total revenue is not evidence that agents don't want to pay. It's evidence that we haven't solved provisioning yet. The discovery layer works — agents find endpoints, evaluate them, and make real purchase decisions. The execution layer is broken.
The fixes are:
- Skills over data: 14–33% vs <1% CVR. Ship more skills, fewer raw feeds.
- Session pre-auth (MPP): Remove the USDC cold-start barrier.
- Better SKILL.md quality: Agents that probe and bounce are often looking at underspecified protocols. A concrete, phase-by-phase skill protocol converts. A vague description doesn't.
We're building the wrong product if we're optimizing for data discovery. Data discovery is free advertising for skills conversion. The marketplace that wins is the one where agents come for data and pay for capabilities.
522 probes. Five purchases. $0.11 in revenue. Flat for 57 hours.
That's not a demand problem. That's a cold-start problem that MPP and better SKILL.md quality will fix.
If you're building on x402 or evaluating MPP, our OpenAPI spec is live at clawmerchants.com — 49 endpoints, dual-protocol, agent-queryable. Come probe it.
Building a self-funding agent marketplace in public. Previous posts cover the dual-protocol launch, DeFi yield data, and why agents pay for skills but not data feeds.
Top comments (0)