A hosted catalogue that turns agent skills from raw GitHub repos into versioned, licensed, signed packages: publish a skill with a manifest and semver, serve it through a signed install endpoint that works across Claude Code, Cursor, Codex and Pi, track installs and updates per version, and enforce paid or seat-limited access for private and commercial skills. The npm registry for the skills economy that October built.

Market Panorama

Competitive Intelligence

Users & Demand

Business Model

Market Sizing

Tech Solution & Build Plan

Go-to-Market Strategy

Risk Assessment

Action Recommendations

### Market definition and scope

The skill catalogue sits between authors who publish capability and developers who install it. It turns a raw repository into a package: a manifest declaring name, version, harness compatibility and requested permissions; a signed artifact; a semver history with changelogs; an install endpoint that emits the correct format for each harness; telemetry that tells the author who installed what version; and an entitlement layer that can enforce private catalogues, seats and paid distribution.

The boundary matters because the adjacent space is crowded in the wrong places. It is not a package registry in the npm sense — a skill is not a library a program imports, it is an instruction bundle a model executes, which means provenance and permissions carry different weight. It is not a harness runtime, not a model host, and not a marketplace of models. Its unit of value is the question an author asks after the first thousand stars: who is using this, which version, and did it break?

October answered the category-size question definitively. Five skill-shaped repositories crossed 4,000 stars within thirty days of creation, led by anatomy at 24,237. In the same window, google-research shipped harness2 (22,206) and openai/math (17,286) — institutional actors following the same distribution pattern as individuals. That combination of a 24,000-star individual project and a 22,000-star research-lab project in the same month is what a category looks like when it arrives.

### Key trend signals with data citations

The distribution unit shifted from a skill to a catalogue. In September the leading repositories were a personal `.agents` directory and a methodology framework. In October they are multi-skill products: replica-skill ships eleven skills for reverse-engineering, rebuilding and debugging applications (7,210 stars), and openfilm packages an entire video capability as an installable unit (4,344 stars).

Commercial intent arrived with the catalogues. heydamir/unslop (6,690 stars) is a single Claude skill that turns vibecoded UI into designed interfaces — a capability someone would pay for, published as a free repository with no licence enforcement and no way to charge. openfilm is the same shape at a larger scope.

Institutional participation removed the last doubt. When google-research and openai both publish skill-shaped repositories in the same month, the format is no longer an experiment by power users; it is how capability is being distributed.

### Why now

Three conditions converged. Scale crossed the threshold where authors notice the absence of analytics — an author with 24,000 stars and no install count is a provoked person, which is the ideal first customer. Commercial intent created a hard requirement for licences and entitlements that a repository cannot satisfy. And cross-harness installation became normal, which means a single-harness marketplace cannot serve the population.

The structural reason the position stays open is more interesting. Every harness vendor wants an ecosystem and none can be neutral, so a vendor registry will always be scoped to one harness. Framework bundles achieve coherence by locking users into a methodology. npm could technically host skills but has no concept of permissions a model executes, so its incentive to build signing and entitlement for this case is weak.

That leaves a third party whose only asset is neutrality and whose only route to legitimacy is author endorsement.

### Market maturity

Early, and the shape of the early market is specific. There is no schema for skill metadata, no convention for declaring requested permissions, and no expectation among authors that a skill should be signed. In a mature registry market those are settled; here the first registry to publish a manifest format people adopt effectively defines it.

That is the strategic opening and the reason a solo founder can compete: the incumbent advantage in registries is standard-setting, and standard-setting favours whoever ships a usable format to an audience first.

### What this means for a solo founder

The product is a manifest format plus author relationships. Build the format in week one, publish it openly, and spend the rest of the first month persuading fifty authors to migrate.

wheresryan22/anatomy reached 24,237 stars within its first month as a single agent skill for isometric technical illustrations.

Skills scale past 24,000 stars in month one

replica-skill ships eleven skills for cloning and debugging apps (7,210 stars); the unit is becoming a catalogue with a brand.

Skill catalogues, not single skills

google-research/harness2 at 22,206 stars and openai/math at 17,286 show institutional investment in the same distribution pattern.

Research labs publish skills

heydamir/unslop (6,690 stars) and openfilm/openfilm (4,344) package agent capability as a product, not a utility.

Agent Skill Catalogue and Entitlement Server