Agent Skills Directory

What is it

An Agent Skill is a packaged, reusable unit of capability that an AI agent can load and execute — think of it as an npm package, but for agent behavior instead of code libraries. A skill pack typically bundles a prompt template, tool definitions, example inputs/outputs, and guardrails, written in a portable format (increasingly the SKILL.md convention popularized by Anthropic's Claude Skills, plus JavaScript/Shell glue for tool execution). The mattpocock/skills repo and the "taste-skill" project are early public examples: versioned, shareable packs that teach an agent how to do a specific job — write a changelog, review a PR, format a spreadsheet — rather than just giving it raw tools.

The business significance is distribution. Once skills are reusable artifacts, they need a registry: discovery, versioning, trust signals, licensing, and payment. That's the same gap npm, the VS Code Marketplace, and the Shopify App Store each filled for their ecosystems. A "skills directory" is the layer that turns scattered GitHub gists into a market. Whoever owns discovery owns the default install path for every agent workflow — a position with real pricing power and strong network effects.

Why now

Three things converged in late 2025 and early 2026 to make this moment. First, Anthropic shipped Agent Skills as a first-class primitive in Claude, and OpenAI and others followed with tool/skill packaging conventions — meaning the format is stabilizing, which is the precondition for any registry. Second, MCP solved the "how does an agent reach a tool" problem, so the remaining pain is "which skill should it use, and can I trust it?" Third, the developer community hit escape velocity: mattpocock/skills and taste-skill are getting organic GitHub traction, and Chinese developer communities (the Juejin piece on skill design) are already debating skill design patterns — a strong signal that practitioners see skills as craft, not throwaway prompts.

Crucially, this is not last year's moment because agents were too unreliable to trust third-party skills, and not next year's because big vendors will likely ship first-party marketplaces once the format wars settle. The window for an independent, vendor-neutral directory is roughly the next 9-12 months. After that, distribution consolidates around whichever platform has the most agent runtime share.

Market Evidence

The raw signal is thin but directionally clean: 2 independent sources (GitHub and Juejin), 3 total mentions, 100% growth rate, first seen 2026-11-06, stage classified as nascent. A 100% growth rate off a tiny base is statistically noisy — three mentions becoming six is not a market. But the composition of the signal matters more than the count. GitHub repos indicate builders shipping real artifacts; a Juejin design essay indicates the community is moving from "here's a prompt" to "here's how to architect a skill," which is the intellectual precursor to a marketplace.

Compare to early signals for other registries: the VS Code extension ecosystem looked similarly sparse in 2015 before exploding. The honest read is real but early demand — a genuine unmet need (discovery and trust for skills) with a market that hasn't formed yet. The opportunity score of 0/100 and demand score of 0/100 reflect that no one has quantified willingness to pay. That's both the risk and the opening: you're not competing for existing spend, you're creating a category. Treat this as a "build the picks and shovels before the gold rush" bet, not a validated market.

Who's Behind It

The visible drivers are individual builders and small communities, not enterprises. mattpocock/skills is the most legible public artifact — Matt Pocock has a large developer-education audience, which gives skill packs a distribution channel they otherwise lack. The "taste-skill" project signals a design-forward faction treating skill quality as a differentiator. On the Chinese side, Juejin (掘金) is the discussion hub, with practitioners publishing skill-design methodology.

The "whales" to watch are the agent runtime vendors: Anthropic (Claude Skills), OpenAI (custom GPTs/tools), and the MCP ecosystem maintainers. None has shipped a full third-party skills marketplace yet, but each has the runtime share to do so. Microsoft/GitHub is the sleeper — GitHub already hosts the repos and could bolt a registry onto Copilot. Your competitive dynamic is therefore asymmetric: you're a small independent building a neutral directory while platform owners sit on the demand side. Your edge is neutrality and curation speed; their edge is default distribution.

TAM & Market Size

Buyers fall into three tiers. Tier 1: individual developers and indie hackers building agent workflows — they'll pay $5-15/month for a curated, versioned skill library, but volume is the game. Tier 2: small SaaS teams (5-50 engineers) standardizing agent behavior across their product — budget tolerance $50-300/month, and they'll pay for private/team skill registries. Tier 3: enterprises needing governance, audit, and private hosting — $1,000+/month, but long sales cycles.

Bottom-up sizing: GitHub has roughly 100M+ developer accounts; if 1% touch agent skills within two years (~1M), and 2-5% of those convert to a $10/month plan, that's $2.4M-$6M ARR from individuals alone. Add team plans and the realistic serviceable market is $50-150M ARR globally within 3-4 years. The demand score of 0/100 is the honest caveat: nobody has proven these buyers will pay for skills specifically rather than getting them free from GitHub. Your TAM thesis rests on curation and trust being worth money — the same bet npm's private registry and the VS Code Marketplace made successfully.

Competitive Landscape

Today the "competition" is mostly absence plus adjacent players. GitHub is where skills live now — free, uncurated, no trust signals, no versioning UX. That's your biggest competitor and your best acquisition channel. Prompt marketplaces (PromptBase, FlowGPT) sell prompts, not executable skill packs, and have weak developer trust. MCP server directories (Smithery, Glama) index tools, not skills — adjacent but different artifact. Agent frameworks (LangChain Hub, CrewAI) host agent configs but are framework-locked and not neutral.

Gaps: no neutral, vendor-agnostic skill registry with versioning, security scanning, and licensing. No trust layer — nobody tells you whether a skill is safe to run. No payment rails for skill authors. Competition score of 0/100 reflects this vacuum.

Big Tech entry risk is high. If Anthropic or GitHub ships a first-party marketplace, you have 9-12 months before default distribution crushes independents. Your differentiation must be neutrality (cross-runtime), curation quality, and author monetization — things platforms are slow to build. Don't compete on breadth; win on trust and the long tail of niche skills platforms ignore.

Business Model

Recommended: freemium SaaS with a marketplace take-rate. Free tier: browse and install public skills, limited to a few installs/month. Pro ($12/month or $99/year): unlimited installs, private skill library, version pinning, security scan reports. Team ($49/month per 5 seats): shared private registry, approval workflows, audit logs. Marketplace: 15-20% take-rate on paid skills, with authors keeping the rest — competitive with the Shopify App Store's 15% and far below legacy app stores.

Why this fits: the value is recurring (skills update, security patches, new packs), so subscription beats one-time. The take-rate creates a flywheel — paying authors attract more skills, which attracts more installs. Pricing rationale: $12/month sits below the $20 psychological threshold for indie devs and above the "why bother" floor; $49/team is cheap enough to expense without approval.

12-month forecast: Conservative $3K MRR (300 Pro users), Base $12K MRR (800 Pro + 40 teams + modest marketplace), Optimistic $35K MRR (2,000 Pro + 120 teams + take-rate). CAC via developer content and GitHub-adjacent SEO: $15-40 per Pro user; payback 2-4 months. Team CAC $200-400, payback ~5 months.

MVP Blueprint

Ship in 5-7 days. Core only: (1) a public skill index that ingests GitHub repos tagged with skill conventions, parses SKILL.md metadata, and renders searchable cards; (2) skill detail pages with install command, version history, and a basic safety badge (does it shell out? does it request network?); (3) a "submit your skill" flow that validates format; (4) email capture and a waitlist for Pro. Cut: payments, private registries, author dashboards, ratings — all v2.

Tech stack: Next.js + Postgres (Supabase) for speed, a GitHub Actions crawler that runs nightly to ingest repos, and Vercel for hosting. Use a simple full-text search (Postgres tsvector) — skip Elasticsearch. Auth via Supabase or Clerk. Total infra cost under $50/month at MVP scale.

Fastest path to launch: build the index first (it's the SEO asset and the proof of demand), then the submit flow, then a landing page with waitlist. Do not build payments until you have 200+ waitlist signups. The MVP's job is to prove people search for and submit skills — that's the signal that justifies the SaaS layer.

Commercial Opportunities

1. Curated skill marketplace for indie devs. Target: solo developers building agent workflows. Expected $5K-15K MRR within 12 months via Pro subscriptions plus 15% take-rate on premium skill packs. Beats building a framework because you monetize discovery, not code — no framework lock-in, and authors do the content work for you.

2. Private team skill registry (B2B). Target: 10-100 person SaaS teams standardizing agent behavior. Sell a hosted private registry with approval workflows and audit logs at $99-299/month. This beats the consumer play on revenue per customer and retention — teams churn slowly once skills are embedded in CI. Expected $10K-40K MRR with 50-150 teams.

3. Skill security scanning API. Target: platforms and enterprises running third-party skills. Sell a scanning API (static analysis for prompt injection, unsafe shell calls, data exfiltration patterns) at usage-based pricing, $0.01-0.05 per scan. This is the picks-and-shovels play that survives even if directories commoditize — trust is the one thing platforms will pay a neutral third party for.

Product Ideas

🥇 SkillHub — "npm for agent skills." A neutral, searchable registry with versioning, safety badges, and one-line install. Target: indie devs and small teams using Claude/OpenAI agents. Why now: format is stabilizing, repos are proliferating, and no neutral index exists. Monetize via Pro ($12/mo) and 15% marketplace take.

🥈 SkillLint — a CI-integrated linter and security scanner for skill packs. One-line value prop: "Catch prompt injection and unsafe tool calls before your agent ships." Target: teams shipping agents to production. Why now: as skills get installed from third parties, supply-chain risk becomes real, and no scanner exists. Sell as a $29-99/month SaaS or usage API.

🥉 SkillForge — a no-code skill builder that turns a team's internal SOPs into versioned skill packs. Value prop: "Turn your runbooks into agent skills in an afternoon." Target: ops and support teams, not just engineers. Why now: the Juejin design discussions show demand for skill authoring methodology, and non-engineers will pay to package expertise without writing code. Price at $39/month per editor.

SEO Opportunity

Search interest is nascent but climbing — "agent skills," "claude skills," and "SKILL.md" are the head terms, all low volume today with high future intent. Long-tail targets: "how to write an agent skill," "claude skill examples," "agent skill marketplace," "mcp vs agent skills," "skill.md format." SEO difficulty is 0/100 — essentially uncontested, which is rare and time-limited. Content strategy: publish the definitive SKILL.md reference and a "best agent skills" roundup now; own the vocabulary before platforms do. Every skill detail page is also a long-tail landing page.

Risk Assessment

The thesis breaks if skills never become a portable standard — if each vendor keeps a proprietary format, a neutral directory has nothing to index. Top risks: (1) Platform absorption — Anthropic or GitHub ships a first-party marketplace and default distribution wins; (2) Format fragmentation — no cross-vendor convention emerges, killing the "neutral registry" premise; (3) Monetization failure — developers refuse to pay for what GitHub gives free, and the take-rate never covers costs.

Validate cheaply: build the index-only MVP in a week, then measure whether people search for skills and submit their own. If 200+ submissions and 1,000+ monthly organic visitors arrive within 60 days, the demand is real. Walk away if, after 90 days, submission volume is flat and waitlist conversion is under 2% — that means the market is either too early or structurally unwilling to pay. Also walk if a major platform ships a competing registry with better defaults before you have 500 users.

Action Plan

Today: register the domain, spin up a Next.js landing page, and manually curate 30 skills from GitHub into a static index. Post it to Hacker News and relevant Discord/Slack communities to gauge reaction.

Week 1: ship the searchable index and submit flow. Instrument analytics on search queries and submissions — these are your two validation metrics. Start publishing SKILL.md reference content for SEO.

Month 1: hit 100 submitted skills and 500 waitlist emails. Launch Pro tier with Stripe. Begin the security-scanning feature as the paid differentiator.

Month 3: reach $3K MRR or 1,000 active users, whichever comes first. If signal confirms, build the team registry (B2B) and approach 20 small SaaS teams directly. If signal is weak, pivot to SkillLint as a standalone security tool — the trust layer is the most defensible piece regardless of directory outcomes.

Related Terms

MCP (Model Context Protocol) — the tool-connection standard; skills sit on top of MCP, so MCP adoption directly expands the skills market. AI Agent marketplaces — platform-owned directories (OpenAI GPT Store, Claude integrations) that are the main competitive threat and validation signal. Prompt engineering tooling — the precursor category; skill design is its professionalized successor, and its users are your early adopters.