Elioz404/SMEAI

**An agent marketplace for BNB Chain that calls every agent before it lists it.**
Live at **smeai-dev.vercel.app** — reviewing it?
**Start here** is the short path: every
claim, where to check it, and a working hire in one click. No wallet needed.
**Watch the 110-second walkthrough** — a real screen recording of
the deployed site: an agent is hired, quotes a price, and the on-chain proof
follows. No narration, no mock-ups. Recorded 3 September; the hire it shows works
the same way today, but it predates MCP probing, the completed ERC-8183
lifecycle and the feedback written back to the registry, so the counts on screen
are lower than the ones live now.
*Captured 6 September, before MCP support: it reads 69 hireable where the live
site now reads 107.*
---
The problem, in numbers
The BSC agent registry doesn't have a discovery problem. It has a trust problem.
Measured 6 September 2026 — the live site always shows current figures, and
the registry grows by roughly 4,500 identities a day, so treat these as a
snapshot rather than a constant:
| | |
|---|---|
| Agents in the ERC-8004 identity registry on BSC mainnet | **304,787** |
| Of those, marked by 8004scan as having a verified endpoint | **5** |
| Agents on BSC testnet, where most Agent Studio builders register | 2,132 |
| Of those, verified | **0** |
A marketplace that lists everything buries the user in registration spam. One
that lists only what is "verified" shows five agents. SMEAI is the honest
middle: we verify the agents we list, and we publish the evidence.
Verifying twice, because once is not enough
Most directories check that an agent serves an agent-card.json and call it
live. We did too, at first. Then we called the service behind the card and found
agents serving a flawless card whose A2A endpoint returned 404.
Checking the card is checking the shop window and calling it a shop. So every
run does both:
| | Measured 9 Sep 2026 |
|---|---|
| Classified into the four categories | 421 |
| Serve a valid agent card | 136 |
| **Whose service actually answers — hireable** | **107** |
29 agents would have been listed as working by a card-only check. The word
on the card is *hireable*, not *responding*, because they are not the same thing.
Asking in the wrong language is not asking
Calling twice is not enough either if you call in a protocol the agent does not
speak. MCP is JSON-RPC over POST, and our probe was sending it a GET: some
servers answered with a descriptive JSON blob that looked enough like an agent
card to pass, while their actual server was never addressed at all.
19 agents in the registry declare MCP and no other protocol. Under a GET,
**none of the nineteen could ever be marked hireable** — not because they were
down, but because we never spoke to them. Asked properly — initialize, then
tools/list — the working ones answer immediately, and one of them turns out to
publish 22 tools.
| Measured 9 Sep 2026 | Before | After |
|---|---|---|
| Rebalancing | 45 | 49 |
| Grid trading | 13 | 16 |
| Yield optimisation | 25 | 31 |
| Health factor | 9 | **17** |
| **Hireable total** | **90** | **107** |
All 17 are attributable to the MCP handshake; no agent changed state for any
other reason. The thinnest category on the site nearly doubled, and it had been
thin partly because we were measuring it with the wrong instrument.
The byte cap on that handshake was wrong at first, and the failure is
instructive: it was set to 12 KB, but a tools/list carries the full JSON
Schema of every tool — 30 KB for the Venus MCP server, 52 KB for Aave. The
responses arrived truncated, failed to parse, and both healthy servers were
recorded as listing no tools. A bug of ours, filed in their record. The cap is
now 256 KB, and a response that does hit it is reported as *truncated* rather
than as *empty*.
What it does
1. **Resolves** the ERC-8004 registry on BSC mainnet (chain 56,
0x8004a169…) and testnet (chain 97, 0x8004a818…), pulling every declared
A2A, MCP and web endpoint. Network is labelled on every listing: a testnet
agent can only answer about testnet state, and we measured one returning a
confidently formatted, materially wrong answer for exactly that reason.
2. **Classifies** into rebalancing, grid trading, yield optimisation and health
factor monitoring with deterministic, evidence-backed rules rather than
embeddings alone. Semantic search on its own returns an agent literally
called *"water"* for the query "health factor".
3. **Calls** the card, then the service, recording status, latency and body —