How this base is built and checked
Last verified
An entry gets into the base only after its licence is read from the repository file, its scripts are scanned without being run, and every model fact is matched to the vendor's own page with a date. Tested means a named person ran it on a sample brief and recorded what broke.
Inclusion criteria
The base lists 36 models and 78 repositories. A further 11 repositories were reviewed and withheld, and 3 models are named without facts because their vendor pages could not be read to the required depth. Skills, MCP servers, tools, awesome lists and single skills all pass through the same gate.
An entry gets its own page only if a computed check passes. The check lives in lib/ai-hub.ts and is never set by hand. An entry that fails stays as a table row with an outbound link.
| Gate | Requirement |
|---|---|
| Read depth | The repository was read in full, not judged from its README alone. |
| Licence | Class A or B, with the licence file read at a pinned commit. |
| Safety | A safety scan is recorded for the entry. |
| Own value | An operator note of 80 to 200 words, plus a test note or two use cases with inputs. |
| Sources | Two or more, all checked within 60 days. |
| Uniqueness | At least 40% of words unique against the nearest sibling page, checked by script. |
| Length | 450 words or more of page-specific text. |
| Status | Not archived and not class C. |
The index is capped at 30 entry pages for the first 6 months. Filters never create URLs, and compare pages stay noindex until their own gate passes.
Licence classes
Class A means a permissive licence was read in the repository file, so the entry may be mirrored with attribution. Class B means link and describe only. Class C means the entry is not listed and is only counted. The current split is 49 in class A, 29 in class B and 11 in class C.
The licence is read at repo level, and a single skill can carry another one. The anthropics/skills repository has no repo-level file and one LICENSE.txt per skill, and some of those are source-available. Each skill therefore needs its own reading before anything is mirrored.
- Class B, no licence file: with no LICENSE file all rights are reserved. Examples are openai/skills, vercel-labs/agent-skills and remotion-dev/skills. These get a link only.
- Class B, source-available: pipeboard-co/meta-ads-mcp (BUSL-1.1) and PostPlusAI/postplus-skills (PolyForm Shield 1.0.0) are readable but are not open source.
- Class B, mixed files: modelcontextprotocol/servers states a transition from MIT to Apache-2.0, so licences differ between files.
- Class B, vendor-bound: some MIT packs only work with a paid service. The hub page discloses that dependency.
- Class C reasons: clone farms, no licence combined with thin content, archived repositories, signpost repositories and red flags in the safety review.
MIT and Apache-2.0 still require attribution. A mirror page shows the repository URL, the licence name, the commit or date mirrored, and what changed. Apache-2.0 also needs the NOTICE file kept. Trademarks and vendor logos fall outside MIT, so logos are not mirrored. An awesome list can have CC0 text while every linked item keeps its own licence, so linking is fine and copying linked content is not.
Safety review
Scripts are listed and scanned for patterns. None has been executed. For the rest of the base, the review used file lists and file heads fetched over the GitHub API, and nothing was downloaded to disk. Regex hits such as subprocess, exec(, urllib, fetch and API_KEY mostly come from repository tooling, evals and build steps.
The scan-skill.py script has been run on the four entries with a full read, at their pinned commits on 15 November 2026, and every flagged high-severity line was read. It runs on a checkout of the pinned commit, and it must run before any other item is hosted or gets its own page. Packs with large script layers need a manual read before anyone installs them, for example aaron-he-zhu/aaron-marketing-skills (174 scripts), AgriciDaniel/claude-ads (129 scripts) and every-app/open-seo (1,116 scripts).
Read before write is the rule for ad accounts. Entries that can publish or change campaigns, such as pipeboard and klosk/adloop, need a preview and approval step, and the entry states the scope each one asks for. Entries where data leaves the machine, such as Firecrawl, Apify and DataForSEO, say so in the entry. Install counts are not quoted for repositories where they run far above stars, because those counts can be inflated or tied to one automated onboarding.
Freshness policy
Each data type has its own re-check interval and a rule for what the page shows when the interval is missed. Model prices are the shortest, since vendors change them on known dates. The dates now known are 30 November 2026, 1 December 2026 and 1 January 2027.
| Data | Re-verify every | Stale rule |
|---|---|---|
| Model price and status | 30 days, plus known vendor dates | Over 45 days the price is hidden and the row shows re-checking. |
| Skill and MCP licence, last push, archived flag | 30 days, by script over the GitHub API | A licence change sets the page to non-indexable until it is re-read. |
| Skill content and test note | 90 days or a new upstream release | Over 120 days the test note is hidden and the page is noindex. |
| Job and guide pages | 60 days | A reviewed date is shown. Over 90 days a review task opens. |
The script seo/ai-hub/check-fresh.mjs needs no API key. It reads lib/ai-base.ts, queries GitHub for licence, pushed_at and archived status, lists stale model rows and writes a dated report. The owner or an agent then re-reads the vendor pages for flagged rows and updates the checked date. The updated date moves only when copy changes.
What tested means
A test note records a run on a sample brief, what the output looked like and what broke. Reading a README is not a test. A model entry needs a same-brief test with the output shown before it can claim a comparison, and no single best model is named without one.
The four entry notes are currently drafts until the site owner has read them. Draft test notes carry no reviewer name. Once reviewed, a note shows who ran it and when.
Conflicts of interest
Three entries are first-party packs and are marked as first party. There are no paid placements and no affiliate links. Vendor-bound packs are listed with their paid-service dependency stated. The base does not describe any third-party repository as endorsed by, or affiliated with, this site.
Corrections contact
Send corrections to the site owner, Vlad Yasinski. Include the entry, the field you dispute and a source page for the right value. A corrected row gets a new checked date and a line in the changelog.
Dataset licence
The file /ai/base.json is licensed CC BY 4.0. You may reuse it, including commercially, if you credit roimobi.com. Licences of the listed repositories stay with their authors and are not changed by the dataset licence.
Re-check changelog
Each re-check gets a dated line here with the data type, the entries touched and the reason, for example a licence change or a price update after a vendor date. A re-check that finds nothing to change is still logged with its date. First line: 15 November 2026, base built from vendor pages and repositories, four entries read in full.
Sources
Each source shows the date it was last checked.
- GitHub github.com/aaron-he-zhu/aaron-marketing-skills
- GitHub github.com/anthropics/skills
- GitHub github.com/every-app/open-seo
- GitHub github.com/kLOsk/adloop
- GitHub github.com/modelcontextprotocol/servers
- GitHub github.com/pipeboard-co/meta-ads-mcp
- GitHub github.com/PostPlusAI/postplus-skills
- GitHub github.com/remotion-dev/skills
- GitHub github.com/vercel-labs/agent-skills
- GitHub AgriciDaniel/claude-ads: LICENSE
- GitHub AgriciDaniel/claude-ads: README.md
- GitHub AgriciDaniel/claude-ads
Found an error, or want an entry checked or added? Send a note.
Contact