Is anyone using this?
An observatory that measures everyone else publishes its own engagement too — the good numbers and the zeros. This page is the demand evidence for the go/no-go decision on making Robin Saige more than an instrument, kept in public so the decision can be audited later.
The funnel, in order of commitment
Reading it left to right: found us → installed the client → read the evidence → asked us to watch a server → reported back after a call. Each step is a bigger commitment than the last.
The endpoint, day by day
Both charts share ONE scale, on purpose — the gap between them is the story. Handshakes without calls = crawlers and the curious. The day calls grow while handshakes hold steady is the day agents start depending on the verdicts.
view as table
| day | tool calls | handshakes |
|---|---|---|
| 2026-08-15 | 240 | 401 |
| 2026-08-14 | 60 | 430 |
| 2026-08-13 | 3 | 39 |
What gets asked
Which tools carry the traffic. should_i_use and check_server are the product; if they lead, agents come to vet — not to browse.
Which software connects
Every MCP handshake declares its client's software name — the one identity dimension we keep (it names a program, never a person). Diversity here is stronger demand evidence than raw volume: many client kinds means the audience found us independently. robinsaige-selftest is us.
Who visits the site
Browsers are people; agents are programs reading pages directly; crawlers index and never engage; self is our own tooling and the operator's browser. Only browsers + agents count as engagement anywhere on this page (258 self requests excluded this week).
Which robots read us
Crawler visits by named robot. The AI answer engines (gptbot · claudebot · perplexitybot · google-extended) are the ones to watch: their reading is the prerequisite for their citing — and citations are a demand-test criterion.
Distinct visitors, day by day
Counted with a daily-salted hash — the salt is destroyed every UTC midnight, so the same person is a different, unlinkable hash each day. We can count Tuesday's visitors; nobody, including us, can tell if they came back Wednesday. That trade is deliberate.
Installs, day by day reported
From pypistats.org — a third-party counter we cannot re-derive; it lags 1–2 days and excludes known mirrors. Day-one spikes usually include package scanners, so again: the trend is the signal, not the level.
What this page will never hold: IP addresses, full user-agent strings, search terms, or tool-call arguments. What it holds, disclosed in full (telemetry v2, 2026-08-15): path + status + a coarse visitor class · referer host · edge country code · a daily-salted visitor hash whose salt is destroyed at rotation (irreversible after the day ends, unlinkable across days) · the MCP client's self-declared software name. IPs and cookies are examined in memory to classify and hash, then discarded. Operator? mark this browser as self. Page telemetry began 2026-08-15; MCP logging 2026-08-13 — history shorter than that isn't missing, it never existed.