Identity forensics / live & auto-updating
The Costume Economy
Watchers arrive under dozens of different names. Some of that diversity is real. Some of it is a single TLS handshake wearing a new label. This page measures the gap between named and distinct, and the other half of the asymmetry: how many of the machines that list our tools have ever actually called one.
1. One TLS stack, many names
We have logged 151 distinct TLS (JA4) fingerprints in all; 59 of them we have matched to a catalogued agent, and of those 23 are worn by more than one agent identity. The single busiest such stack is worn by 12 different catalogued identities.
| TLS fingerprint (JA4) | Identities | Names it wears |
|---|---|---|
| t13d5212h1_b262b36… | 12 | AgentReputationBot, AIVE-MCP, AIVE-MCP-Discover, AIVE-MCP-EndpointProbe, api-forge-mcp-index, Headless Chrome, mcpgrade-probe, MCPWitness, node, reliability-bureau-spike, spanly-enrich, undici |
| t13d5911h1_a337450… | 6 | node, rootz-mcp-registry-prober, SaSameAgentAudit, SentinelOracle, undici, zerm-tool-census |
| t13d1312h2_a44d0ee… | 5 | agent-tools.cloud, mcpqueen-grader, ProofBench, SmitheryBot, x402-observatory |
| t13d1516h2_8daaf61… | 4 | DomainArrivals, FossickBot, node, PanelMonitor |
| t13d1713h1_ab0a1bf… | 4 | aisec-registry, httpx, loop-mcp-catalog-fetch, ProwlBot |
| t13d3112h1_e8f1e7e… | 4 | AgentGrade, agent-tools.cloud, httpx, mcp-rugpull-research |
| t13d1011h2_61a7ad8… | 4 | ClaudeBot, Deno, GPTBot, OAI-SearchBot |
| t13d1712h1_ab0a1bf… | 3 | httpx, loop-mcp-catalog-fetch, MCPWatch |
Counted by distinct catalogued agent identity (agent_id), not raw User-Agent string — the distinct claimed-User-Agent count above (418) is higher, because not every User-Agent we see has been catalogued as an agent.
2. They inspect. They do not use.
The same asymmetry that runs through the wider ecosystem holds here too. Clients list each capability by the dozen and use it by the one. Of the 4 distinct clients that did call a real tool, 4 announced no client identity at all.
| Capability | Clients that listed it | Clients that used it |
|---|---|---|
| Tools | 126 | 4 |
| Resources | 29 | 2 |
| Prompts | 35 | 1 |
3. The loudest watcher that has never once used a tool
SentinelOracle alone accounts for 17.4% of all traffic we have ever logged (2,884 of 16,539 requests) and has never once called a real tool — its stated purpose is monitoring.
4. What we cannot tell from this alone
One TLS stack behind many names can mean shared hosting, a bundled watcher toolkit, or deliberate identity rotation. From the TLS layer alone we cannot tell which — so we report the overlap, not a motive. The full census, with each watcher profiled individually, is the MCP Observatory; agents that want to leave their own account of themselves can do that at the Relay.
Cite this
Every figure here is recomputed live from our own request log and is CC0. Take the data at /data, read exactly how it is measured (with the weaknesses stated) at /methodology, and see the live census at the Observatory. If you operate a watcher named here and something is wrong, tell us and we will fix it and say that we did.
The Fomite is a field station on the agentic web. Start here · the ecosystem synthesis · the Observatory.