State of MCP: A Longitudinal Measurement Dataset of the Model Context Protocol Ecosystem (v2026-10-31)

State of MCP is a longitudinal, firsthand measurement of the public Model Context Protocol (MCP) server ecosystem by Major Labs (majorlabs.co). Weekly scans began in June 2026 and the series cannot be backfilled. Discovery unions four sources (GitHub topic and organization search, the official MCP registry, Glama, Smithery), deduplicated by repository. Analysis is read-only static inspection of public source code; no server is ever run or probed. The security and identity heuristics are tuned for precision over recall, so their prevalence figures are lower bounds. This version, v2026-10-31, was cut from the October 31, 2026 sweep. It catalogs 3,745 repositories (3,227 in v2026-Q3) with public GitHub metadata and derived transport and manifest signals, records 202 discovery runs, and carries the longitudinal series through 13 snapshots from June 18 to October 31, 2026. No aggregate snapshot was recorded between August 29 and September 27. Security sweep, 3,150 servers: 94 in the High tier, 1,121 Elevated, 1,935 Low. Server-side request surface appears in 35.1 percent of scanned servers, code execution patterns in 4.0 percent, command injection patterns in 1.8 percent, unsafe deserialization in 0.5 percent and hardcoded credentials in 0.4 percent. Identity sweep, 3,091 servers: 949 take sensitive actions such as payments, email, file or shell access, and 728 of those show no authentication check in static analysis, an identity gap of 76.7 percent (78.6 percent of 733 in v2026-Q3). Supply-chain sweep, 2,732 servers: 2,226 (81.5 percent) declare at least one dependency whose declared version, or the lower bound of its range, matches a published advisory of any severity. This is not a count of exploitable servers. The weekly series reads 72 to 75 percent from June to late August and 34 percent on August 29, when advisory lookups ran short while dependency counts stayed complete. It reads 68 to 69 percent on September 27 and October 3, after the number of servers with manifests nearly doubled, and 78 to 81 percent from October 10. v2026-Q3 was cut on September 9, inside that shortfall, so its 36 percent is an undercount, not a method change. The README's Known issues section has the detail. The aggregate history now carries the risky-pattern share, the identity gap and the advisory-match share for every snapshot, computed from the stored per-server snapshot rows. Its risky share and high-surface counts match the weekly figures published on majorlabs.co. The high-risk cohort remediation series (47 servers, frozen June 10, 2026) is carried over unchanged from v2026-Q3. Per-repository security, identity and supply-chain findings are never published. They go to maintainers through coordinated disclosure, and the dataset holds population metadata and aggregates only. MANIFEST.json lists every file with its SHA-256 checksum; README.md is the data dictionary. Data: CC BY 4.0. Scanner code: MIT, github.com/major-matters/mcp-scanner. Landing page and live series: majorlabs.co/data.