OpenAgentSkill

What It Is (confidence: medium)

OpenAgentSkill (Leon-Drq/openagentskill, 200+ stars at capture) is a skill registry + recommendation API built for agent-native discovery: “Let your AI agent find, compare, and install the right reusable skill automatically.” Its big bet: ordinary directories are for people to browse; OpenAgentSkill is built so an agent can discover, compare, and install skills programmatically. Indexed 20,842 skills, 860K+ downloads, 104 agent surfaces.

Architecture (confidence: medium)

Four layers between intent and install:

1. Intent capture — a human or upstream agent describes the job in natural language (Task · Agent · Context).
2. Recommendation engine — skills ranked by workflow fit, quality, freshness, stars, and audit signals (Fit · Quality · Freshness).
3. Skill trust profile — each candidate gets readiness notes, install commands, review prompts, and risk summary (Risk · Install · Evidence).
4. Agent install path — returns the next action an agent can safely execute (Codex · Claude Code · Cursor · Install).

Agent-facing API, e.g. POST /api/agent/resolve → GET /api/agent/skills/crawl4ai → GET /api/skills/crawl4ai/install?format=text. The registry response for “analyze stock news” returns recommended_skill, install_command, why_recommended (matches research workflow, strong Trust Score, audit warnings included), and risk_summary (safety: “review before install”, notes: network access / verify sources).

Positioning (confidence: medium)

The captured comparison table positions OpenAgentSkill against three alternatives:

Feature	OpenAgentSkill	skills.sh	agentskills.io	Native docs (OpenAI Codex)
Primary job	Recommend, compare, install from one registry	Browse and install reusable agent skills	Define the open skill format and learning path	Explain skills inside each native platform
Agent-facing API	Yes — task-to-skill recommendations	Directory and install workflow	Spec and documentation first	Platform-specific APIs and docs
Cross-agent	Codex, Claude Code, Cursor, MCP, custom runtimes	Open agent skills ecosystem	Open format for extending agents	Best for the vendor platform
Trust/audit	Stars, quality score, readiness notes, install review	Directory metadata	Metadata guidance in SKILL.md	Native platform controls

The point stated is complementarity, not replacement: OpenAgentSkill focuses on the registry + recommendation layer agents can call.

Implications: OpenAgentSkill pushes the skill lifecycle further than static directories — it adds a decision layer (which skill, from intent) and an audit layer (trust before install) that no static catalog provides. That is the missing piece for the skills-directory-gate pattern to scale past a handful of curated skills: when the manifest is 20k entries, routing-by-scanning fails and intent→recommendation is the workable alternative. Its trust profile is a governance answer to skill-authoring-best-practices’s emphasis on discoverability — the registry adds the safety metadata the authoring doc doesn’t cover.

Open Questions

• Trust/fit scores are self-declared registry signals; no public evidence ties them to measured task success.
• Indexed-skill count (20,842) vs skills.re’s ~7,380 — different inclusion criteria; not directly comparable.
• The recommendation engine’s ranking weights are unspecified; an agent cannot audit why fit=96 vs 92 beyond the listed bullets.

Related

skills-re — the agent-routed registry; OpenAgentSkill routes on task instead skills-directory-gate — the runtime pattern this recommendation API is the scaled-up version of skill-authoring-best-practices — authoring rules for the SKILL.md artifacts the registry distributes skill-reducer — measured evidence on description/body compressibility, relevant to trust metadata design ponytail — a real portable skill the registry layer could recommend for coding tasks

Sources

• raw/external/openagentskill-com-6cae86ca.md