Research
Original MCP ecosystem research
The data behind Vouch’s scores, published in full. Each is a dated snapshot against a stated corpus — see the methodology for how the underlying scores are produced.
- The MCP census — August 2026 (9.8% sample)
An early, partial look at the registry: the transport split, a credential-class breakdown, and an estimate of how many MCP tools are probeable — from a 9.8% sample.
Published 2026-08-22 · 9.8% systematic sample of the remote-transport corpus (1,347 of 13,708 servers)
- The deep-probe study — MCP tool behaviour
A partial replication of the Digital Applied MCP reliability study, run against 129 real, live-probed tools.
Published 2026-08-23 · Full behavioural probe battery against 164 curated read-only tools; 129 completed, 118 published a score
- The MCP census — September 2026
The whole registry, not a sample: the transport split, credential-class breakdown, and the counted number of MCP tools Vouch can actually probe.
Published 2026-09-02 · Full census of 24,813 deduplicated servers; every one of the 13,781 remote-transport servers probed (no sampling)