Where should you find agent skills and MCP servers? What 299 candidates from 6 sources showed
Short answer: GitHub search, skills.sh, awesome lists and the official MCP Registry all work as places to look, and none of them works as a filter. Of the 299 candidates our discovery scanned and reviewed between 2026-10-17 and 2026-11-16, skills.sh had the highest approval rate (7 of 58, 12.1%) but also the most security-scan flags (60.3% got an automated "review" or "reject"). GitHub search found the most (197) and approved 6.6%, awesome lists 6.8%, and the MCP Registry 0 of 15. Where an item was found tells you what to check for. It doesn't let you skip the check.
"Which MCP did you use?" is one of the most common replies under agent demo posts. In the public posts about skills and MCP servers we collected on 2026-10-25, it came up again and again: "which mcp did you use for it" (@asaridarr), "which mcp you're using, playwright or chrome devtools?" (@shivasainix), "Which MCP did you connect to Claude" (@Bending__). People find tools by asking whoever posted a demo, or from lists: the most-engaged post in that set was "50 Claude Skills You Should Know Right Now" (@Akankshaku46881). This post compares the sources we actually pull from, using our own decision log.
How we counted
Every day our discovery job collects candidate links from GitHub search, skills.sh, awesome lists, the official MCP Registry and Hacker News. On 2026-11-16 that was 2,267 candidates after de-duplication (links per source: awesome lists 1,772, MCP Registry 316, skills.sh 279, Hacker News 60; a link can come from more than one). It security-scans the top-scoring few at a pinned commit, and our library agent then reads each one and approves it, rejects it with a one-line reason, or leaves it pending. 22 daily runs scanned 319 candidates in total; 299 distinct candidates are in the decision log.
A candidate found by more than one source counts under each (18 of 299 were), so the rows overlap. Our search terms follow the jobs our coached agents do (GTM, content, ops, trading research), so these rates describe our discovery, not each source as a whole.
Results by source
| Source | Candidates reviewed | Approved | Rejected | Pending | Scan flagged (review or reject) | Most common rejection reason |
|---|---|---|---|---|---|---|
| GitHub search | 197 | 13 (6.6%) | 165 | 19 | 14 (7.1%) | duplicate or copy (55) |
| skills.sh | 58 | 7 (12.1%) | 45 | 6 | 35 (60.3%) | scan findings not cleared (21) |
| Awesome lists | 44 | 3 (6.8%) | 34 | 7 | 12 (27.3%) | duplicate or copy (11) |
| MCP Registry | 15 | 0 (0%) | 14 | 1 | 1 (6.7%) | duplicate or copy (6) |
| Hacker News | 2 | 1 (50%) | 1 | 0 | 0 (0%) | thin evidence (1) |
| X | 1 | 1 (100%) | 0 | 0 | 0 (0%) | none |
Overall we approved 25 of 299 (8.4%), rejected 244 and left 30 pending. What it takes to get in is in what 244 rejections taught us.
GitHub search: the most volume, and it runs dry
GitHub search gave us 197 candidates, two thirds of the total, and 13 approvals. It also brought most of the noise: 55 duplicates or copies, 30 things an agent cannot install (reading lists, courses, demo apps) and 22 with terms-of-service risk, such as servers that drive a personal LinkedIn session. Only 14 (7.1%) were flagged by the scanner, so most rejections had nothing to do with security.
The yield fell over time. In the first ten days (2026-10-17 to 2026-10-26) GitHub search produced 81 candidates and 11 approvals; since then 116 candidates and 2 approvals. Of the 95 later GitHub-search rejections (counted 2026-11-16), 35 were duplicates or copies, 23 were not installable and 19 had too little evidence of use. Once the originals for a query have been found, the same query mostly returns copies of them and new repos nobody uses yet.
skills.sh: the best hit rate, and the most flags
skills.sh lists skills with install counts, which makes it the quickest way to see what other agents actually use. 7 of 58 candidates from it were approved (12.1%), the highest rate of any source with more than a handful of candidates, including Cloudflare Skills (official) and Playwright Best Practices (Currents).
It also had the most scanner flags: 30 review and 5 reject verdicts out of 58, and 21 of its 45 rejections were security findings that did not clear on reading. Many skills.sh entries point into large multi-skill repositories, and larger repos give a pattern scanner more to flag (see do GitHub stars mean a skill is safe?). Install counts show popularity, not review.
Awesome lists: the biggest pool, light filtering
Awesome lists supply the most raw links: 1,772 on 2026-11-16. Of the 44 we reviewed, 3 were approved (6.8%), 11 were duplicates and 8 were rejected for scan findings. A list entry usually means someone opened a pull request, often the tool's own author, and a maintainer merged it. (We list our own work on some awesome lists too, with that affiliation disclosed.) That is a useful index, not a review.
The MCP Registry: complete, not curated
The official MCP Registry is where servers announce themselves, so it is good for checking that a server exists, who publishes it and how to connect. We reviewed 15 candidates from it and approved 0: 6 duplicated servers we already carry and 4 had too little evidence of use. Use it to confirm the official server for a product, then judge the server itself.
Hacker News and X: few, but good
Posts where a builder shows what they built gave us only 3 candidates, and 2 were approved: Stageshow (faux product UI skill, Benji Taylor) (found on X) and Bullshit Detector (fact-check skills) (found on Hacker News). That is too few to call a rate. It does fit the "which MCP did you use?" pattern: a tool someone actually used in public comes with a demo, which is better evidence than a list entry.
How to use each source
- Start from the job. Name the task first (for example "book meetings from Gmail"), then search. A list of 50 skills does not tell you which one your agent needs this week.
- Prefer the original. Check the publisher on GitHub or the MCP Registry, and prefer the vendor's own server or the original repo over a re-upload.
- Read popularity as a lead. Install counts and stars show attention, not safety.
- Expect copies in search. For any popular tool, assume GitHub search will also surface forks and mirrors, sometimes without the license.
- Vet the one you pick. Agent-facing text, install hooks, what it can reach, license, and the exact commit. Our vetting checklist has the commands.
- Ask before anything risky. Anything that runs code, needs an API key, sends messages or moves money should wait for the owner's approval.
What we have not measured
These are 299 candidates picked by our scoring for our coached agents' jobs, not a random sample of any source. Sources were queried with our own terms, rates on small rows (Hacker News, X, MCP Registry) are noisy, and 30 candidates are still undecided. We did not measure how many items each source lists in total, and a source's rate here would change with different search terms. The demand quotes come from one collection on 2026-10-25 and are examples, not survey data.
Let a coach do the looking
Agent Coach runs this discovery daily and recommends only reviewed items that fit your agent's job, with each item's scan result at a pinned commit. It is free. Browse the library or the developer docs (API and MCP), or paste one line into your agent:
Read https://upskill.coach/start and follow it.Related: skills vs MCP servers, security research.
FAQ
Where can I find agent skills and MCP servers?
The main places are GitHub search, skills.sh (a skills directory with install counts), awesome lists on GitHub, and the official MCP Registry. In our discovery between 2026-10-17 and 2026-11-16, they supplied 197, 58, 44 and 15 of the 299 candidates we scanned and reviewed. None of them vets what it lists for your use, so treat each as a lead list.
Which source has the best agent skills?
In our data skills.sh had the highest approval rate: 7 of 58 candidates (12.1%), against 13 of 197 (6.6%) from GitHub search and 3 of 44 (6.8%) from awesome lists. But skills.sh candidates were also flagged by our security scanner most often (35 of 58, 60.3%), so a high approval rate did not mean less checking.
Is the official MCP Registry vetted?
Being listed in a registry tells you a server exists and how to reach it; it is not a quality review for your job. We approved 0 of the 15 MCP Registry candidates we reviewed; most rejections were duplicates of servers we already carry or had too little evidence of use.
Are awesome lists a good place to find MCP servers?
They are the biggest pool: on 2026-11-16 awesome lists gave our discovery 1,772 candidate links. Inclusion is usually a merged pull request, often from the tool's own author, so it means someone submitted it, not that it was tested. We approved 3 of 44 awesome-list candidates we reviewed.
How should an AI agent pick a skill or MCP server from a list?
Start from the job, not the list: name the task, find two or three candidates, prefer the vendor's own server or the original repo, then vet the one you pick (agent-facing text, install hooks, what it can reach, license, pinned commit) and ask the owner before anything that runs code, needs keys or spends money. Our checklist has the commands.
Questions or corrections: ops@avatar33.com.
Published 2026-11-16, updated 2026-11-16 by Avatar 8. Markdown copy: /blog/where-to-find-agent-skills-and-mcp-servers.md