I let 250 bots crawl my API registry for 48 hours. Here is what they actually asked for

Two days of logs from a machine-facing API registry: which manifest files agents read, what they search for, and the one thing that makes an entry usable by an agent.

I built agentnexus.app, a registry of APIs, MCP servers and CLIs meant to be read by AI agents rather than by people. Then I did the only interesting thing you can do with a machine-facing site: I logged every single machine request and read all of them.

In 48 hours: 1 965 requests, 250 distinct clients, and not one human sign-up. Here is the useful part of that data.

1. Agents don't read your homepage. They read your manifests.

Ranked by hits over 24 hours:

Surface Hits
/mcp 485
individual entry JSON 254
/sitemap.xml 36
/openapi.json 30
/llms.txt 28
/.well-known/agent-card.json 28
/.well-known/oauth-protected-resource 27
/.well-known/mcp.json 22
Nobody asked for HTML. The MCP endpoint took 40 % of all traffic, and the second most-read thing was the per-entry JSON — the machine version of a product page.

Practical takeaway: if you want agents to use your service, the artifacts that matter are an MCP endpoint, an OpenAPI document, a llms.txt, and a .well-known card. Everything else is decoration.

1. A surprising share of traffic is auditors, not consumers

The recurring clients, verbatim from the logs:

mcpbeat/0.1 (+https://mcpbeat.com/bot/; liveness check) 56
AgentTrust-Monitor/1.0 (+https://agenttrust.site/methodology) 46
rokmcp-collector/0.2 (+https://rokmcp.com/bot) 42
AgenstryBot/0.3.0 (+https://agenstry.com/bot) 40
PREA-DiscoveryEngine 30
GolemreachTrustBot/0.1 22
mcp-contract-scanner/0.1 (research; read-only tools/list) 11
Around fifteen of these run on a fixed schedule — PREA hits exactly every two hours, three requests per round, always the same three files. They are not calling my tools. They are checking that I still answer, that my auth declaration hasn't changed, and that my schema still matches what they recorded yesterday.

There is a whole layer of the ecosystem whose product is "this MCP server is alive and behaves". Two of them published a listing about my server without me ever submitting it. One of them sells verified-alive endpoints to agents, per call.

Which means: your uptime and your schema stability are now public, third-party facts. Break your tools/list shape and a dozen registries record the regression within the hour.

1. tools/list is not enough. Agents need a call they can copy.

The registry exposes a discover_capabilities tool: an agent sends a need in plain language ("send a transactional email", "weather forecast") and gets back matching interfaces with endpoint, auth contract, formats, rate limits and reliability history.

The single change that made entries actually usable was adding a tested curl example to each one. Not a template with YOUR_API_KEY — a command I ran, that returned 200, stored as-is. 127 of 155 entries have one. The remaining 28 all require a key or an account, and that is exactly the information an agent needs before it tries.

If you publish one thing for agents today, publish a request that works when pasted, and say plainly whether it needs credentials.

1. Where the traffic comes from

US 612, Belgium 175, Netherlands 127, Finland 120, Germany 76, France 42. Almost entirely datacenters. Two clients look like genuine consumers rather than monitors: a Python process on a 10-minute loop, and a node client spread over 14 machines in 5 countries.

And one line in the logs is a human who put his email address in his own user agent string, which I found weirdly charming.

1. The uncomfortable conclusion

Machine discovery works, and it works without any SEO, without a launch post, without a single backlink from a human. Directories found me, measured me, and published me on their own.

What none of that produced: a single sign-up. Agents read, monitor, score and republish. They do not create accounts and they do not pay. The people who do that still arrive through posts like this one.

So if you're building for agents: build the manifests, they will absolutely find you. But don't confuse being indexed with being adopted.

If you want to poke at it, everything is open without auth:

curl https://agentnexus.app/llms.txt
curl "https://agentnexus.app/api/public/discover?need=send+a+transactional+email"
curl https://agentnexus.app/openapi.json
MCP endpoint: https://agentnexus.app/mcp

I'm curious what your own logs look like — if you run anything machine-facing, what fraction of your traffic turns out to be auditors?