The Liveness Law
24K Findings · Four registries, 204,500 registered agents and services measured.
The cheaper it is to get listed, the more of that registry is dead.
| Registry | Cost to list | Dead |
|---|---|---|
| ERC-8004 on-chain identity | gas only | 85–97% |
| CDP Bazaar | free | 67–79% |
| Glama MCP registry | curation + scoring | 47% unhealthy (their own figure) |
Sort by entry cost and they line up. Free entry selects for abandonment.
We probed CDP Bazaar three times across July rather than once, because one number for a moving population is a moment, not a fact: 22,545 services at 74% dead, then 25,614 at 79%, then 24,583 at ~67%. Three runs, one direction.
What counted as alive: a valid HTTP 402 challenge, or a served x402 manifest. Not a 200. A 200 from a marketing page is the most common way a dead service looks alive, and any count that accepts it is measuring DNS, not commerce.
Two details worth the ink
Of fifteen endpoints listed in registries’ own machine-readable discovery feeds — the interfaces built for agents, not humans — eight returned 404 at the advertised path. The surface designed for machines was wrong more often than not.
And every headline registry count you will read is working software and abandoned software added together, presented as one number.
Buying is harder than finding
In July we bought services off our own shelf to confirm delivery: 16 of 126 were purchasable by a machine at a discoverable address, 8 delivered exactly what they advertised, 0 took payment and returned nothing. $0.054 spent, every transaction reconciled on-chain.
The interesting number is the first one. Almost nothing failed after payment — a machine could not get as far as paying, for want of a findable address or a documented request shape. The friction in this economy sits before the payment, not after it. Most services are fine. Most front doors are not.
What we cannot claim
All four measurements are Western. ModelScope’s MCP plaza renders as a JavaScript app we could not enumerate; Zhihu and India’s NPCI disallow crawlers and we honored that. The mechanism is economic, so we would expect it to hold elsewhere — expecting is not measuring. If you can sample a non-Western registry, check us. A refutation is worth more to us than agreement.
Our own position
We run a directory, and this page is an argument that directories are mostly graveyards. It points at us too: 454 hand-checked entries, small by every number above, and our own delivery check found most of that shelf not machine-purchasable either. Listing with us means a maintainer probed it and it answered correctly at the time of review, and we re-check. That is the whole claim — not an audit, not an uptime guarantee.
If you’re building: list somewhere free and you are shelved among the dead. Consume a registry and presence is not evidence anything works. Probe before you spend.
Method: one throttled probe per host to each advertised endpoint; 49 blocked sources recorded as blocked rather than dropped from the denominator. Corrections by pull request or issue.
— 24K Labs