Skills are the new packages.
SkillDB is where agents go to find them.
The agent skills standard is here. Anthropic shipped it, OpenAI adopted it, and Vercel started a leaderboard. What's missing is the curated, packaged, breadth-first library agents can call from inside an MCP session — not another grab-bag of community repos. That's SkillDB: 4,557+ skills, 394+ packs, 37 domains, one install.
$ claude mcp add skilldb -- npx skilldb-mcp
The model is smart. The knowledge it has access to isn't.
Every developer using an AI coding agent runs the same loop: ask, get a generic answer, Google for domain-specific context, copy-paste the relevant bits back into the chat, re-ask. The bottleneck stopped being model intelligence around Claude 3.5. It's knowledge logistics now.
The current loop
- Agent gets a task: "audit this auth flow for OWASP issues."
- Outputs a generic checklist that misses your stack's real risks.
- You search, open seven tabs, copy the relevant sections back into chat.
- Re-prompt. Output is 10× better.
- Tomorrow, different domain, same loop.
SkillDB: a curated library of agent expertise. Your agent finds it itself.
One MCP server. One install. Your agent now searches 4,557+ practitioner-grade skills, picks the right ones for the task, loads them inside its context budget, and reasons with them. No prompt-stuffing. No RAG pipeline. No retraining.
Agent calls skilldb_search / skilldb_recommend with task context. Returns ranked skills in milliseconds.
skilldb use auto detects the project stack and activates only matching packs — no token waste, no off-task expertise.
Skills land in .skilldb/active/. The agent reads them like any other context — no API roundtrips at runtime.
Three standards collapsed into a window. This is the window.
Twelve months ago, "agent skills" was a term of art. Today it's a published spec with cross-vendor buy-in and a fast-moving distribution layer. Every plumbing question got an answer. The remaining question is who builds the catalog.
$10.9B today. $50B in 2030. Skills are the picks-and-shovels layer.
The agentic AI market is doing the npm-shaped thing it always does: model providers, agent runtimes, and tool ecosystems each grow into a wedge. The skills layer is the part nobody wins by default, and the part every other layer leans on.
The split
One catalog. Three integration paths. Zero retraining.
We meet agents where they are. MCP for Claude Code, Cursor, and anything else that speaks the protocol. A CLI for power users and automation. A typed REST API for everyone building bespoke stacks.
claude mcp add skilldb
10 native tools — search, get, recommend, suggest, list, set_key, my_skills, create_skill, update_skill, purge.
npm i -g skilldb
skilldb search · skilldb add · skilldb use auto. Auto-detects project stack and activates the right packs.
GET /api/v1/skills
Bearer-token auth, batch retrieval, scoped keys. Free metadata; full content on Pro+.
Differentiators
- Curated packs, not a grab-bag.Every pack ships as a unit — install once, get the matching set.
- Context-budget aware.use auto loads only what fits, ranked by stack relevance.
- Breadth past engineering.Writing, screenwriting, design, finance, ops — 37 categories total.
- Skills.md compatible.Same frontmatter Anthropic uses. Drop-in for Claude Skills users.
- Zero runtime calls.Skills cached locally. No API hop on every agent turn.
- Private skills (Studio).Bring-your-own internal standards alongside the public catalog.
From cold start to live catalog in three months.
We've compounded faster than the standards we're catching. Daily skill additions, six automation pipelines that keep distribution warm, and a paying-tier funnel already in front of users.
What we learned
The fastest-growing skills aren't the obvious ones. Screenplay adaptation, novel audit, vibe-coding workflow, and culinary skills got more traffic in March than half the engineering packs. The agent-skills market doesn't end at code — and most of the catalog isn't being built for the long tail.
The honest landscape.
We're not pretending the field is empty. It isn't. We're betting on a different shape of product than what's out there.
| Player | Shape | Where they win | Where we win |
|---|---|---|---|
| skills.sh (Vercel) | Open community directory + leaderboard | Discoverability of ad-hoc community repos | Curated packs, breadth, paid full-content tiers |
| anthropics/skills | Reference repo + canonical examples | Authority, integration with Claude Skills product | Volume, multi-vendor, non-engineering domains |
| Prompt libraries | Static text repos / Notion / spreadsheets | Familiar, free, easy to start | Agent-discoverable, structured, versioned |
| Custom RAG | Per-org pipelines and vector stores | Private knowledge, deep integration | Setup in seconds, no infra, multi-domain by default |
| Fine-tuning | Model-baked expertise | Latency, cost-per-token at scale | Updatable in minutes, portable across models |
Read: skills.sh tracks ~91k community-submitted skills; SkillDB ships 4,557+ that we've actually curated, packaged, and gated. Different product. Same standard.
Free to discover. Paid to use at speed.
Metadata and copy-paste are free forever. Programmatic access — bulk packs, API quota, MCP full content, private skills — is metered. Same shape as npm Pro / GitHub Pro / Sentry. Pricing already live; gating enforced via Stripe.
3 CLI/day · metadata · bookmarks · community.
Unlimited CLI · bulk packs · 10K API/mo · auto-sync.
100K API/mo · private skills · webhooks · priority builds.
Enterprise is the obvious next wedge: private skill registries with the same MCP/CLI shape, billed per seat or per pack. Three discovery calls in flight; not booking yet.
One operator. Builds in public. Ships on Fridays.
Former Google Senior Staff Engineer (AI Platform). Founding engineer at Vercel (#14). Fraud detection at Stripe. Forward deployed at Palantir. Stanford M.S. CS (AI & HCI). Princeton B.S. CS & Math.
Built SkillDB because he was burning 6–9 hours a week as a copy-paste API for his own agents. Three months later: 4,557+ skills, 394+ packs, an admin platform, an outreach engine, and an automation pipeline that runs the marketing loop while he writes more skills.
“The bottleneck was never the AI. It was the knowledge. Models have been smart enough for a while — they lacked structured, practitioner-level expertise in a format they could use.”
Hiring plan post-seed: 2 senior engineers (skill quality + API scale), one DevRel, a curation team for domain packs.
$2M seed. 18 months to category default.
Enough runway to take SkillDB from "the curated catalog people install" to "the catalog enterprise teams pay for." Use of funds is allocated against milestones, not headcount theatre.
| Allocation | % | Outcome |
|---|---|---|
| Engineering (2 senior) | 40% | Skill quality system, multi-tenant API, enterprise registries. |
| Curation team | 25% | Scale to 15k+ skills; domain experts for specialized packs. |
| Go-to-market / DevRel | 20% | Conferences, partnerships with Claude Code / Cursor / Codex. |
| Infrastructure | 10% | Cloud, CDN, monitoring, security audit, compliance. |
| Operations | 5% | Legal, accounting, the boring necessary stuff. |
18-month milestones
- →15,000+ skills · 50+ categories
- →10,000+ MAU on CLI / MCP
- →$50K+ MRR from Pro & Studio
- →3+ enterprise contracts (private registries)
- →Skill marketplace with community contributions
- →Default install in two of: Claude Code · Cursor · Codex docs
Want the deck, the demo, or both?
Investors, partners, and design partners — same inbox. We're also happy to send a 10-slide PDF if your partnership wants the file.
Sources: AI agents market sizing — IDC & Gartner (2026); Anthropic Agent Skills launch (Oct 2025); Vercel skills.sh launch (Jan 2026); Anthropic MCP standard donation. Stats live as of build time and pulled from the SkillDB index.