Loading
What agents buy from other agents, settled on chain.Note 23
ERC-8004 agent cards held on chain · 14,441 declare a service · read 2026-10-03
10,506 agents publish something a machine could call. 17 of their 1,290 owners have ever been paid on a rail we read.
Offers, never calls. An MCP request is HTTP between two parties and nobody outside them can count it, so nothing here stands in for usage. What a card declares is public, and it is a claim in exactly the sense a registration is: free to write, and checkable against whether the host answers and whether the owner has ever earned anything.
The last figure is not evidence that the tools are unused. It says the money for them does not move on a rail we read, which is a different sentence, and the one an agent charging a card or invoicing a company would produce.
A card saying “MCP” without an endpoint is a capability nobody can reach. Counted apart rather than folded in, because the two are different claims.
| Declared as | Agents | Entries | With an endpoint |
|---|---|---|---|
| A website | 10,645 | 10,663 | 10,649100% |
| MCP servercallable | 10,290 | 10,317 | 10,310100% |
| Something else, named | 10,260 | 10,933 | 9,90591% |
| OASF schema | 9,978 | 9,978 | 9,968100% |
| A2A skillcallable | 9,594 | 9,598 | 9,596100% |
| A wallet address | 9,210 | 9,278 | 9,277100% |
| A social handle | 355 | 838 | 838 |
1,140 declarations over 618 distinct names. Counted by agents declaring each, so a card listing one twice is one agent offering it.
33,031 declarations over 309 names, and they are mostly taxonomy paths: an agent filing itself under a category rather than naming a function anybody can invoke. Kept out of the tool count for that reason.
One MCP handshake per declared endpoint, identified as us: the same initialize and tools/list any client sends on connecting. Nothing is invoked and no credential is sent, so a server that demands one is recorded as private rather than as broken. Endpoints are asked once however many registrations name them, which matters here: one host carries 472 of them.
No tool is ever invoked: calling somebody else’s tool to score its host is their cost and our claim. So a listed tool is a tool the server says it has, and nothing on this page claims a task would succeed.
A registration can carry its card or point at one. These are the ones we fetched: https and ipfs only, private addresses refused, each fetch kept with its outcome and the document’s hash. A card read is a claim read; whether the host answers and whether the owner was ever paid are the checks elsewhere on this page.
Most declared: web (14,316) · A2A (5,893) · agentCard (5,603) · trackRecord (5,603) · Termix Platform (5,161) · MCP (1,581) · q402 (801) · RESTAP (405)
Outcomes: 28,754 read, 1,147 not-json, 994 http-429, 291 http-404, 126 unreachable, 46 unsupported, 25 private-target, 8 http-502, 6 http-503, 6 too-large, 5 http-402, 5 http-530, 2 http-403, 1 http-401, 1 http-500. A refusal on rate is asked again the next day; a card we could not reach is not a card with nothing in it.
Every declared MCP server is asked for its tools and every declared A2A endpoint for its agent card, and the answer is kept by day. A capability is counted as changed only where both days read the list: a list we could not read is an unknown, not an empty one.
68 of 891 asked endpoints served on the latest day (no-handshake 721, serving 68, unreachable 63, error 23, gated 16).
Of 891 endpoints seen on an earlier day too: 1 stopped serving, 1 started, 0 changed their tools (+0, −0), 0 changed version.
Seen on two days or more: 893. Served on every one of them: 66. On some and not others: 5.
| Host | What changed | Detail |
|---|---|---|
| mcp.hergertsynthora.com | stopped serving | now unreachable |
| namewhisper.ai | started serving | was unreachable |
349 of 891 asked endpoints served on the latest day (not-a-card 517, serving 349, unreachable 14, gated 11).
Of endpoints seen on an earlier day too: stopped serving, started, changed their skills (, ), changed version.
152 of 195,901 on-chain agent cards say which model they run on: 12 in a field made for it, the rest in a sentence such as “powered by”. Another 281 only mention a model, which read one by one is mostly a topic they cover (“GPT-4o vs Claude”), so they are not counted as running on it.
What they say they run on, by family: claude 109, gpt 12, qwen 12, gemini 7, grok 6, llama 4, deepseek 2.
| Model | Agents | Input, per M tokens |
|---|---|---|
| anthropic/claude-opus-4.6 | 5 | $5.00 |
| anthropic/claude-sonnet-4.6 | 3 | $3.00 |
| openai/gpt-5.2 | 3 | $1.75 |
| qwen/qwen3-coder | 2 | $0.300 |
| anthropic/claude-sonnet-4.5 | 2 | $3.00 |
| openai/gpt-4 | 2 | $30.0 |
| openai/gpt-5.4 | 2 | $2.50 |
| x402 priced resource | 214 | 263 | 263100% |
|---|
The hosts behind MCP and A2A endpoints, from the daily health read rather than a fresh probe.
“Answered HTTP, not MCP” is not a claim that the endpoint is not a server. It covers a card pointing at a web page, and it also covers a real server on the older HTTP+SSE transport, which this probe does not speak, counted above rather than left as a caveat. The only endpoints not asked are the ones whose host NAME does not resolve: a root that refuses or times out says nothing about whether the MCP path answers, and three of the servers above were found exactly there.
What the private ones ask for: 10 refused without naming a scheme, 6 oauth. A server that demands credentials is a server that exists.
652 across 68 servers, 407 of them distinct.
Identical tool sets are one deployment under several names as often as they are several products, and nothing in a registration distinguishes the two. Counted rather than judged.
The second number is a FLOOR and not a duration. The longest of them has been down for at least 24.1 days, which is as far back as this watch reaches. A surface that died a year ago and one that died the day we started look identical from here, and saying otherwise would be inventing the difference. Only the first number is an observed death.
A declaration is one line in a registry; the thing behind it sits on somebody’s machine.
The same question of the servers that completed the handshake. The two lists barely overlap.
| Server | Tools | Protocol | Endpoint |
|---|---|---|---|
| topaz-agent | 70 | 2024-11-05 | https://agents.topazdex.com/mcp |
| nw-terminal | 44 | 2025-06-18 | https://namewhisper.ai/mcp |
| xona-agent | 39 | 2025-06-18 | https://api.xona-agent.com/mcp |
| whitemagic | 24 | 2024-11-05 | https://mcp.whitemagic.dev/mcp |
| cortensor-mcp-http | 19 | 2024-11-05 | https://router0-t1a.cortensor.app/mcp/messages |
| cortensor-mcp-http | 19 | 2024-11-05 | https://router1-m1.cortensor.app/mcp/messages |
| pretium-mcp | 18 | 2025-06-18 | https://mcp.pretium.africa/mcp |
| bnbagent-seller | 17 | 2025-06-18 | http://43.160.214.54:8000/mcp |
| clash-of-coins | 16 | 2025-06-18 | https://mpp.clashofcoins.com/mcp.json |
| clash-of-coins | 16 | 2025-06-18 | https://x402.clashofcoins.com/mcp.json |
Seen on two days or more: 891. Served on every one of them: 346. On some and not others: 10.
| Host | What changed | Detail |
|---|---|---|
| explainer-agent.onrender.com | stopped serving | now not-a-card |
| www.whitemagic.dev | version | 9.3.1 → 9.3.4 |
Served is not used. A server listing a tool says the tool can be selected; nothing here calls one, and no figure here is a count of calls, customers or revenue.
| openai/gpt-5 |
|---|
| 1 |
| $1.25 |
| anthropic/claude-opus-4.7 | 1 | $5.00 |
|---|
| anthropic/claude-opus-4.5 | 1 | $5.00 |
|---|
| openai/gpt-4o | 1 | $2.50 |
|---|
A card naming a model is a claim, not a call. A model is matched to the catalogue only when the name is exact; “built with Claude” names a family and carries no price.
Of 2,101 declared MCP and A2A endpoints, 1,163 sit on a hosting platform we can name from the domain. Of the 68 that served on the latest day, 20 do.
| Platform | Serving | Declared |
|---|---|---|
| Vercel | 16 | 62 |
| Railway | 2 | 1,024 |
| Cloudflare | 1 | 24 |
| Google Cloud | 1 | 9 |
| AWS | 0 | 15 |
| Render | 0 | 13 |
| Fly.io | 0 | 10 |
| GitHub Pages | 0 | 4 |
| ngrok tunnel | 0 | 1 |
| Netlify | 0 | 1 |
An endpoint on a domain of its own names no platform and is not guessed at. The platform is where the service is served from, which is a dependency whether or not the agent says so.