Daily snapshots of the MCP Registry: first numbers

The official MCP Registry shows what’s listed now, but it doesn’t keep a history. Since 6 November 2026 I’ve been taking a full snapshot every day, plus read-only tools/list probes of public hosted servers (initialize + list tools only — never calling a tool). This is a baseline from the 7 November 2026, 06:36 UTC snapshot (mcp-shelf-archive snapshot-2026-11-07). Two days of history so far, so nothing here is a trend yet. What’s on the shelf 57,993 entries: 57,090 active and 903 deprecated. Overnight from 6→7 November: 158 added, 3 removed. 19,109 active entries have a latest-version date in October 2026; 9,295 more are dated 1–7 November (that’s the date of each entry’s latest version, not when it was first listed). Who publishes 36,027 publishers. 91% list exactly one server. Ten publishers account for about 15% of active entries; the single largest GitHub-namespaced account lists 2,366. Source links and downloads 34.7% of active entries give no source repository. Of 858 linked GitHub repositories checked, 104 (12%) no longer exist; 55% of the rest have zero stars. Across 4,643 servers whose npm packages we checked: median 81 downloads/week; the top 1% take about 90% of all downloads in that sample. Tool lists (hosted, public HTTPS only) Asked 550 hosted servers; 388 answered openly; 123 required a login. Median server exposes 9 tools; the largest in the sample, 165. Of 6,621 tools listed, 31% carry no annotations at all (no readOnly / destructive hints). Of 152 servers already seen the day before, 16 changed tool definitions overnight — 14 of those 16 were one publisher updating the same shared tools across its servers. Read that as one night’s data, not a rate. Method (short) Registry: public MCP Registry API, latest version of every entry, daily. Tool lists: public, no-login, HTTPS endpoints; rotating sample + fixed watch list; initialize and tools/list only. Downloads / repos: official npm and GitHub APIs, daily rotating samples. No logins, no scraping, no tool calls. Limits Samples rotate and aren’t the whole shelf. Counting entries isn’t judging quality. Corrections welcome: jordan@bot2botlabs.com. Full write-up with the same figures and limits: https://bot2botlabs.com/r/mcp-gh-202611-a