Open research · 2026 edition
State of Agent Skills 2026.
A transparent look at what is actually inside an agent-skill registry: which records look like reusable skills, which are maintained, which declare a license, what can be installed, and how little real outcome evidence the ecosystem still has.
22,001
Indexed candidates
1,000
Analyzed sample
9%
Strict agent skills
55
Agent Proven
Executive finding
Popularity is not readiness.
GitHub adoption is useful, but it does not prove task fit, safe installation, or a successful agent run. This report keeps those signals separate.
- 018.8% of the analyzed sample has strict Agent Skill evidence in its repository name, README, description, or topics.
- 0212.3% describes an agent workflow or agent product without strict reusable-skill evidence.
- 0329.4% is better described as a domain workflow tool or ecosystem dependency than a strict agent skill.
- 042.2% is a list or collection rather than one installable skill.
- 0599.7% was pushed within 180 days of report generation.
- 0681.7% has a declared license that is not marked unknown.
- 07100% reaches the current 72-point high-trust threshold in this quality-ranked sample; Trust remains a metadata signal, not proof of task success.
- 0855 analyzed skills have at least one reported agent outcome, covering 75 outcomes in total.
Readiness signals
What the analyzed sample can support.
- Maintained within 180 days
- 997100%
- Known license
- 81782%
- Install path available
- 1,000100%
- 500+ GitHub stars
- 1,000100%
- 72+ Trust Score
- 1,000100%
- Safe-to-try audit
- 85285%
Domain coverage
Where the candidate supply is concentrated.
Track assignment uses category, task, use-case, and repository text signals. It is a discovery aid, not a claim that every record is production ready.
| Track | Candidates | Share | Maintained | Known license | High trust |
|---|---|---|---|---|---|
| Coding and developer agents | 495 | 49.5% | 493 | 406 | 495 |
| Research and knowledge work | 172 | 17.2% | 172 | 144 | 172 |
| Design and creative production | 127 | 12.7% | 127 | 102 | 127 |
| Data, BI, and analytics | 98 | 9.8% | 98 | 79 | 98 |
| Legal, policy, and compliance | 36 | 3.6% | 36 | 29 | 36 |
| Finance and quant workflows | 34 | 3.4% | 34 | 30 | 34 |
| Marketing and growth automation | 25 | 2.5% | 25 | 17 | 25 |
| Education and tutoring | 8 | 0.8% | 7 | 5 | 8 |
| Presentation and deck workflows | 3 | 0.3% | 3 | 3 | 3 |
| Football and World Cup analytics | 2 | 0.2% | 2 | 2 | 2 |
Evidence shortlist
Leading strict skill candidates.
Ordered by reported outcomes, Trust Score, skill-likeness, and GitHub adoption. This is a shortlist for evaluation, not an automatic-install endorsement.
| Skill | Track | Stars | Trust | Audit | Risk | Outcomes |
|---|---|---|---|---|---|---|
| Frontend Design anthropics/skills | Design and creative production | 171K | 89 | 93 | Safe to try | 7 |
| Vox Director Alisa0808/vox-director | Design and creative production | 1.5K | 82 | 91 | Safe to try | 4 |
| Talos siderolabs/talos | Coding and developer agents | 11K | 88 | 91 | Safe to try | 3 |
| Taste Skill: Anti-Slop Frontend Leonxlnx/taste-skill | Design and creative production | 80K | 92 | 94 | Safe to try | 2 |
| 女娲.Skill alchaincyf/nuwa-skill | Research and knowledge work | 28K | 88 | 91 | Safe to try | 2 |
| Crossplane crossplane/crossplane | Coding and developer agents | 12K | 88 | 91 | Safe to try | 2 |
| Radzen Blazor radzenhq/radzen-blazor | Data, BI, and analytics | 4.3K | 88 | 92 | Safe to try | 2 |
| Carbon carbon-design-system/carbon | Design and creative production | 9.2K | 87 | 91 | Safe to try | 2 |
| Firecrawl firecrawl/firecrawl | Research and knowledge work | 139K | 87 | 90 | Safe to try | 2 |
| Obsidian Second Brain eugeniughelbur/obsidian-second-brain | Research and knowledge work | 3.6K | 85 | 92 | Safe to try | 2 |
| Keystatic Thinkmill/keystatic | Research and knowledge work | 2.1K | 84 | 89 | Safe to try | 2 |
| Longhorn longhorn/longhorn | Coding and developer agents | 7.8K | 84 | 89 | Safe to try | 2 |
| Animation Vocabulary emilkowalski/skills | Design and creative production | 17K | 90 | 93 | Safe to try | 1 |
| WhisperX m-bain/whisperX | Design and creative production | 23K | 90 | 92 | Safe to try | 1 |
| Kedro kedro-org/kedro | Design and creative production | 11K | 90 | 93 | Safe to try | 1 |
| Aaron Marketing Skills aaron-he-zhu/aaron-marketing-skills | Coding and developer agents | 2.6K | 89 | 94 | Safe to try | 1 |
| Echarts apache/echarts | Data, BI, and analytics | 67K | 89 | 91 | Safe to try | 1 |
| Ppt Master hugohe3/ppt-master | Presentation and deck workflows | 37K | 89 | 92 | Needs review | 1 |
| Dashi Ppt Skill chuspeeism/dashi-ppt-skill | Presentation and deck workflows | 5.2K | 88 | 92 | Safe to try | 1 |
| Crawlee apify/crawlee | Research and knowledge work | 24K | 88 | 91 | Safe to try | 1 |
| Maxun getmaxun/maxun | Research and knowledge work | 16K | 88 | 92 | Safe to try | 1 |
| Appsmith appsmithorg/appsmith | Coding and developer agents | 40K | 88 | 91 | Safe to try | 1 |
| Crawl4AI unclecode/crawl4ai | Research and knowledge work | 73K | 88 | 91 | Safe to try | 1 |
| Scrapegraph AI ScrapeGraphAI/Scrapegraph-ai | Research and knowledge work | 27K | 88 | 91 | Safe to try | 1 |
| InsForge InsForge/InsForge | Coding and developer agents | 12K | 88 | 92 | Safe to try | 1 |
Methodology
Readable before repeatable.
Published Jul 10, 2026. Data refreshed Aug 26, 2026.
Population and sample
All public records with ai_review_approved=true in the OpenAgentSkill registry. MCP-only records are excluded by the registry read layer.
The highest-ranked 1,000 approved candidates by the current quality ordering. The sample is a product audit, not a random statistical sample.
Inclusion and measurement
- •Public GitHub repository or repository URL is available
- •Registry AI review is approved
- •MCP-only server records are excluded
- •Skill-likeness is measured separately from repository popularity
- •Trust and audit scores use public repository metadata and declared install information
Limitations
- •GitHub stars measure repository adoption, not agent task success
- •Candidate listings are not equivalent to verified maintainer claims
- •A public metadata audit is not a substitute for source review or sandbox execution
- •Agent Proven evidence remains sparse until agents report real outcomes
- •Repository metadata can change after this report is generated
Cite this report
OpenAgentSkill. “State of Agent Skills 2026.” Published July 10, 2026. Dataset and methodology available at https://www.openagentskill.com/reports/state-of-agent-skills-2026.