ERC-8004 gave on-chain agents a registry. What it did not do — could not do — is tell you what any of those agents can actually do. The registry is an index of identities; the capabilities live off-chain, behind each agent's own endpoints. For the last several weeks we've been doing the unglamorous work of visiting those endpoints, one by one, across three chains, and recording what we find: the Model Context Protocol (MCP) tools they expose, the OpenAPI methods they publish, the AgentCards they serve, and who owns them.
The result is the largest structured view of the ERC-8004 agent economy we know of. As of this writing we have — agents indexed and — enriched profiles on file. The single most interesting number, though, is this one: we have captured — live MCP tool descriptions from — agents that run a reachable MCP server. That is a direct, machine-read inventory of what these agents are building — not what their marketing says, not what their on-chain metadata claims, but the tools they will actually hand to a model at runtime.
Most agents are quiet. The ones that aren't are worth watching.
The first thing the data makes obvious is that the agent economy follows a brutal power law. Of the tens of thousands of registered agents, only — publish a live MCP server, and only — publish a parseable OpenAPI inventory. The overwhelming majority register an identity and never expose a working capability surface we can reach. That is not a knock on the standard — it is what early ecosystems look like. It does mean that the agents which do publish live tooling are a small, legible, high-signal set.
Among MCP publishers, the top of the table is genuinely interesting. As of 1 June 2026, the most tool-rich agents we see are EruditePay Facilitator and Delx (60 MCP tools each, both on Base), followed by Syenite (46) and, on Ethereum, Name Whisper (41). These are not toys — an agent exposing 40 to 60 distinct tools is describing a real product surface: payment facilitation, naming, data lookups, on-chain actions. You can see the current ranking, live, on our leaderboards.
The OpenAPI side tells a slightly different story. A cluster of agents — MolTrust, Paygent Price Oracle, EmblemAI, Agoragentic — sit at the top with very large method counts (our probe caps the inventory it records, and these agents hit the cap). High OpenAPI method counts tend to indicate a service that was built API-first and later wrapped as an agent, rather than designed around MCP from the start. Both patterns are valid; together they tell you the ecosystem is being populated from two directions at once — MCP-native agents and API-native services putting on an agent face.
The long tail is mostly factories and test rigs.
If you go looking for the rest of the distribution, you find what we found in earlier analysis of ownership: the long tail is dominated by batch deployments. Several of the highest MCP-tool agents outside the very top are members of an obvious minted series — a run of near-identical "FakeRover-Berlin" agents (#44322 through #44351 and neighbors), each carrying the same 42-tool manifest. Elsewhere you find frank test deployments — names like "ERC8004-Test-Agent" and "DualReg-Test-v2" sitting on otherwise production-looking infrastructure.
This matters for anyone trying to reason about the size of the agent economy. The headline agent count is real, but a large share of it is factory-minted clusters under a single owner and developers kicking the tires. We track this directly: across the index there are — distinct owner wallets behind all those agents. A meaningful fraction of the agent population collapses into a much smaller number of actual operators once you deduplicate by owner — a single owner in our data holds thousands of agents. When you read "tens of thousands of agents," read it as "tens of thousands of registrations, a few thousand owners, and a few hundred genuinely active capability surfaces."
Identity is the scarcest signal — and the most valuable.
Of those — owners, only — resolve to a human-readable identity through ENS, Basenames, or Farcaster. That is a small percentage, and it is exactly why owner identity is one of the things buyers pay us for: it is genuinely hard to come by. Most owners are bare contract addresses. We classify them where we can — as of now we have flagged — owner contracts as verified on their block explorer, and identified — owners as plain externally-owned accounts rather than contracts. The gap between "an address minted this agent" and "this named team operates this agent" is the gap our enrichment layer exists to close.
Three chains, three very different shapes.
The agent economy is not evenly distributed. As of 1 June 2026 the index holds roughly — agents on BNB Chain, — on Base, and — on Ethereum mainnet. BNB leads on raw registration volume — a lot of that is the factory-cluster pattern described above — while Base is where the most capable, named, MCP-publishing agents tend to concentrate. Ethereum mainnet is the smallest by count, which is unsurprising given gas costs; the agents that do register there skew toward higher-value, longer-lived deployments. If you want to understand "the agent economy," you actually need to understand three different economies with different cost structures and different cultures.
Reachability: the silent filter.
A capability you can't reach isn't a capability. Beyond MCP and OpenAPI, we probe each agent's host for a parseable well-known AgentCard and check basic host health — is the endpoint reachable, is its TLS certificate valid and unexpired. We currently see — agents serving a parseable well-known card. The rest fail the most basic test of being a live service: a request to their advertised endpoint does not come back with something we can use. Host health turns out to be one of the more practical signals in the whole dataset — it is the difference between an agent that exists on paper and one you could actually integrate with this afternoon.
What this tells us about where ERC-8004 is going.
Three things stand out. First, MCP is winning the capability layer: when an agent wants to describe what it can do in a way another agent can consume, it increasingly does so as MCP tools, and the richest agents we track are MCP-native. Second, the economy is far more concentrated than the registration counts suggest — a few hundred live capability surfaces and a few thousand real owners sit underneath a much larger pile of factory mints and test rigs. Third, the scarce, valuable signals are exactly the ones that require leaving the chain: live tool inventories, resolved owner identity, host health. The registry tells you an agent exists. Everything that tells you whether it matters lives at the endpoint.
We'll publish this report on a regular cadence as the numbers move. Every figure above is live from our
public index — you can read the raw aggregates yourself at
GET /v1/public/stats (free, no key), browse the rankings on the
leaderboards, or pull a full enriched profile for any agent through
the paid API.