The MCP ecosystem grew a third in a quarter, fastest where it is least secured
Eighteen weekly scans, May 31 to August 31. The server population is up 31 percent, and the fastest-growing slice is the remote-hosted servers that carry the widest attack surface. The abandonware share held flat. This is a population read, and it is clean.
The population counts here are re-measured every week and are current. The deep security, identity, and supply-chain sweeps are disclosure-sensitive and run manually, not on the weekly job, so those pillars carry a Q2 baseline and are being re-based this quarter. This report reports only the population movement, which is clean. No security-posture delta is claimed.
By the numbers
Window May 31 to August 31, 2026, 18 weekly scans. Every figure is firsthand and checkable at majorlabs.co/security.
Steady, not spiky
The count everyone quotes jumps around because catalogues auto-index anything MCP-shaped. The maintained population we scan firsthand tells a calmer story: a straight line up and to the right, roughly 42 net new servers a week, no single-week surge. This is an ecosystem compounding, not a hype spike.
Firsthand maintained-core count, 18 scans. Endpoint labeled.
The growth went to the riskier surface
Not all growth is equal. A local stdio server that runs on one machine is a smaller target than a remote-hosted server reachable over the network. Over the quarter, remote-hosted servers grew +43 percent against +31 percent for the population as a whole. The part of the ecosystem expanding fastest is the part with the widest attack surface.
Remote is where agents actually reach a server over the wire. It is also the slice growing fastest. Whatever the security baseline turns out to be, it is being applied to a bigger, more exposed surface each week.
Growth is not outrunning rot, but it is not losing to it either
A quarter of the ecosystem is stale, meaning untouched for six months. The notable result is that this share did not move. Stale servers grew 30 percent, active grew 32, and the ratio held at roughly 23.5 percent across all 18 weeks. MCP is adding maintained servers at very nearly the rate it is adding abandoned ones. It is neither maturing nor decaying on net; it is scaling with its composition fixed.
| Slice | May 31 | Aug 31 | Change |
|---|---|---|---|
| Total tracked | 2,413 | 3,166 | +31% |
| Active (≤180d) | 1,841 | 2,421 | +32% |
| Stale (>180d) | 572 | 745 | +30% |
| Remote-hosted | 1,003* | 1,430 | +43% |
| Stale share | 23.7% | 23.5% | flat |
*Remote first cleanly counted 2026-06-06.
Method, and the honest limits
Every population figure is a firsthand, read-only scan of the public MCP ecosystem, one per week, derived from scanner/scans.db. "Active" means a push within 180 days; "stale" means none. "Remote-hosted" counts servers reachable over the network rather than local stdio.
What is deliberately absent: any Q3 movement in the security, identity, or supply-chain pillars. Those deep sweeps are disclosure-sensitive and run manually, so the weekly job carries their last measured value forward unchanged. Reporting that carried-forward number as a quarterly change would be a measurement artifact, not a finding. The next report will re-base those pillars from a fresh deep scan and report their delta honestly from that point.
Cite: Major Labs (2026). The State of MCP — Q3 Delta. majorlabs.co/reports/state-of-mcp-q3.