Monthly numbers on the open-source agent skill and MCP server repositories on GitHub that Agent Skills Hub indexes: how fast they grow, how they fare in our security scan, what licenses they carry and which agents they target.
Repositories indexed
257,330
Skill, MCP and agent tool repos on GitHub
Created in October 2026
29,603
By GitHub creation date
Flagged by the security scan
4.9%
Of 124,557 audited repos
No license
35.0%
Reuse rights unclear
New repositories per month
29,603 new repositories were created in October 2026: 13,224 MCP servers, 4,643 Claude skills and 2,960 Codex skills. Codex skills outnumbered Claude skills among new repositories in 5 of the last 6 months.
By the month each repository was created on GitHub, counting the repositories indexed at snapshot time. Recent months keep filling in as indexing catches up, so every issue keeps its own numbers.
What the catalog holds
36.8% of the catalog are MCP servers; Claude and Codex skills together make up 27.2%.
MCP servers94,726 · 36.8%
Agent tools65,184 · 25.3%
Codex skills36,344 · 14.1%
Claude skills33,688 · 13.1%
Other skills14,446 · 5.6%
Tools8,535 · 3.3%
LLM plugins4,153 · 1.6%
Agent frameworks253 · 0.1%
Other1 · 0.0%
Security scan
124,557 repositories (48.4% of the catalog) have been through the security scan. 4.9% were graded caution or worse, and 361 were graded unsafe. The rest of the catalog is unaudited and is not counted as safe.
🟢 Safe118,447 · 95.1%
🟡 Caution5,748 · 4.6%
🔴 Unsafe361 · 0.3%
⛔ Reject1 · 0.0%
Most common risky patterns
Uses sudo for elevated privileges4,755
Exposes the local network through a tunnel (ngrok and the like)508
An install step in a code block runs a script from an untrusted host378
Reads environment variables in code272
Installs a cron job261
Reads agent memory or identity files227
Spawns subprocesses to run commands187
References several API keys or tokens160
Turns off TLS certificate checks67
Runs code with eval()54
Pipes a downloaded script into a shell (curl | sh)46
Runs Docker in privileged mode46
A pattern hit is a lead, not proof: when we reviewed flagged repositories, 44.5% of hits were text that mentions a pattern (docs, warnings) rather than code that does it. Grades weigh the hits.
What is counted: public GitHub repositories that hold an agent skill, an MCP server or an agent tool, found by our search queries every 8 hours plus curated additions. Counts are repositories, not files: one repository can hold many SKILL.md files, which is why directories that count files report millions.
Growth uses the date each repository was created on GitHub, among those indexed at snapshot time; a repository we have not found yet is missing from every month.
Security grades come from a static scan of README, SKILL.md and scripts. 48.4% of the catalog has been scanned so far.
Types and platforms are inferred from names, topics, READMEs and manifests. They can be wrong for a single repository; the totals are what this page is for.
Snapshot taken 2026-11-01 07:05:52 UTC. A published issue is never changed; corrections go into the next one.
Cite and download
Free to use, also commercially, with credit and a link: CC BY 4.0.
Agent Skills Hub (agentskillshub.top). State of Agent Skills, October 2026. https://agentskillshub.top/stats/2026-10/
