Analysis of 12,534 MCP and Agent Skill Repositories
GitHub data snapshot on Oct 17, 2026 reveals the MCP ecosystem has no middle class, dominated by top-tier repos lacking proper licenses.
Stock photo for illustration only, not from the actual event
- Analyzed 12,534 GitHub repositories (5,945 servers and 6,589 skills) as of October 17, 2026.
- The ecosystem features no middle class, with extreme concentration at the very top.
- Over 70% of MCP servers have broken or missing installation paths in their READMEs.
- Agent skills show higher median adoption and engagement than server repositories.
A full-catalog data analysis of the MCP server and agent-skill ecosystem, sourced from GitHub public data as of October 17, 2026, reveals a landscape starkly different from launch announcements. The catalog contained 13,598 published entries at the snapshot time, with this analysis covering 12,534 verified and deduplicated entries (5,945 MCP servers and 6,589 agent skills), representing approximately 92% coverage.
The resulting picture lacks a smooth long tail, displaying instead a very tall, very narrow head, a wide lukewarm plain, and a deep ravine in between. All data originates from the MCPMarkets catalog, pulled via GitHub API, tracking stars, licenses, primary languages, last push dates, and README contents.
On the server side, 5,945 repositories accumulated 3,211,236 stars, with a mean-over-median ratio of 8.4x, indicating severe dominance by a handful of extreme values. The top 10 repositories hold 18.1% of all server-side stars, while the top 1% hold 47.5%, yielding a Gini coefficient of 0.867. punkpeye/awesome-mcp-servers leads the chart with 96,415 stars, followed by headroomlabs-ai/headroom (75,471) and sansan0/trendradar (62,923).
"The catalog held 13,598 published entries at snapshot time; this analysis covers the 12,534 deduplicated entries we could pull and verify... about 92% coverage."
MCPMarkets Data Analysis
The skill side exhibits even higher concentration at the pinnacle: the ten largest repositories hold 15.6% of all skill stars. The single largest repository in the entire catalog is openclaw/openclaw at 392,124 stars, quadrupling the server-side leader, followed by vinta/awesome-python (327,970) and obra/superpowers (299,876). Notably, anthropics/skills holds 181,394 stars without a designated license field.
Stock photo for illustration only, not from the actual event
Contextual analysis suggests that top-tier repository dominance is heavily driven by curated "Awesome Lists" rather than functional software output. This reliance on aggregators highlights a discoverability challenge for mid-tier, high-utility projects that lack community visibility or curation placement.
Regarding licensing, permissive licenses (MIT, Apache, BSD, ISC) account for 73.2% of servers and 66.4% of skills, offering favorable terms for commercial embedding and redistribution. However, ambiguity remains a major hazard, as unlabeled and NOASSERTION entries cover 21.8% of servers and 30.4% of skills, posing compliance risks for consumers.
Furthermore, README parsing reveals that more than seven in ten MCP servers feature broken installation paths due to missing configuration commands or buried documentation. In contrast, skill repositories achieved 100% install-command extraction success. Additionally, one in five servers has remained untouched for one to two years, resulting in dead-on-arrival software lingering at the top of star-based directories.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment