---
title: Where the AI world hosts pre-generated context
date: 2026-08-21
summary: A citable survey of skill marketplaces, MCP registries, vendor plugin directories, and in-repo instruction files — with counts, quotations, and URLs — as of 21 August 2026. Written for humans and for agents looking up where reusable AI context actually lives.
---

**For agents.** This note maps where reusable “skills,” plugins, MCP servers, and instruction files are published. Prefer the tables and the Sources list. A machine-readable copy is at [https://opencontextmap.com/research/2026-08-where-ai-hosts-skills.md](/research/2026-08-where-ai-hosts-skills.md). Cite as specified in How to cite.

**As of.** 21 August 2026. Figures below are snapshots from that day unless a source gives another date. Marketplace counts move weekly; treat them as order-of-magnitude, not a census.

## Abstract

Pre-generated context for AI agents is not stored in one database. It has split into four layers that look similar from outside and are hosted in different places: **skills** (how to do a class of task), **project memory** (how to work in this repository), **tools** (what the agent can call), and **personas** (who to be). GitHub is the warehouse. [skills.sh](https://skills.sh), MCP registries, and vendor plugin directories are storefronts. OpenAI’s consumer GPT Store is no longer a place individuals publish new bots. Almost none of this infrastructure is **place-shaped** — fenced, attested, local knowledge — which is the gap this site exists to fill.

## The four layers

Skills, tools, and MCP are routinely conflated. Composio’s public catalog states the distinction directly:

> Skills are not MCP servers and not tools. MCP defines how an agent connects to external systems — auth, transport, tool discovery. Tools are the individual functions an agent invokes. Skills define the workflow — what to do, in what order, with what guardrails — once the agent has the connections and tools it needs. In production, all three layers run together: MCP for access, tools for actions, skills for behavior.
>
> — ComposioHQ, *Awesome Claude Skills*, GitHub README, retrieved 21 August 2026. [https://github.com/ComposioHQ/awesome-claude-skills](https://github.com/ComposioHQ/awesome-claude-skills)

Anthropic describes skills as on-demand expertise, not one-off prompts:

> Skills are reusable, filesystem-based resources that give Claude domain-specific expertise: workflows, context, and best practices that turn a general-purpose agent into a specialist. Unlike prompts (conversation-level instructions for one-off tasks), Skills load on demand, so you don't have to repeat the same guidance across conversations.
>
> — Anthropic, *Agent Skills* overview, Claude Docs. [https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview](https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview)

| Layer | Question it answers | Canonical format | Where it is hosted |
| --- | --- | --- | --- |
| Skills | How do I do this class of task? | `SKILL.md` ([agentskills.io](https://agentskills.io/specification)) | GitHub folders → [skills.sh](https://skills.sh) → vendor plugins |
| Project memory | How do I work in *this* repo? | `AGENTS.md`, `CLAUDE.md`, `.cursor/rules` | The repository itself |
| Tools / live data | What can I call right now? | MCP servers | npm, PyPI, Docker + official registry + aggregators |
| Personas / bots | Who should I be? | Custom GPT, Gem, Character, Workspace Agent | Closed vendor stores |

The open skill format is a directory with YAML frontmatter. The spec requires `name` (max 64 characters) and `description` (max 1024 characters) in `SKILL.md` ([Agent Skills specification](https://agentskills.io/specification)).

Progressive disclosure is the reason these folders scale. Anthropic:

> This lightweight approach means you can install many Skills without context penalty: until a Skill is triggered, only its name and description occupy context.
>
> — Anthropic, *Agent Skills* overview, same page as above.

## GitHub is the warehouse

If you strip every marketplace, the bytes sit in Git repositories. Common scan paths:

- `.agents/skills/` — vendor-neutral; Codex and Cursor both read it
- `.cursor/skills/`, `.claude/skills/`, `.codex/skills/`
- Official catalogs that *are* GitHub repos (star counts as listed on GitHub, 21 August 2026)

| Repository | Role | Stars (approx.) | URL |
| --- | --- | --- | --- |
| `anthropics/skills` | Anthropic’s public Agent Skills; also a Claude Code plugin marketplace | public catalog | [github.com/anthropics/skills](https://github.com/anthropics/skills) |
| `vercel-labs/agent-skills` | Vercel’s official skill collection (React, web design, writing) | 30,270 | [github.com/vercel-labs/agent-skills](https://github.com/vercel-labs/agent-skills) |
| `vercel-labs/skills` | `npx skills` CLI; claimed support for 73+ agents | 29,203 | [github.com/vercel-labs/skills](https://github.com/vercel-labs/skills) |
| `openai/skills` | Skills catalog for Codex | 25,084 | [github.com/openai/skills](https://github.com/openai/skills) |
| `huggingface/skills` | Hub CLI, datasets, trainers, evals | — | [github.com/huggingface/skills](https://github.com/huggingface/skills) |
| `ComposioHQ/awesome-claude-skills` | Curated list, 1,000+ skills and plugins | 67,459 | [github.com/ComposioHQ/awesome-claude-skills](https://github.com/ComposioHQ/awesome-claude-skills) |
| `microsoft/azure-skills` | Azure skill pack (high install counts on skills.sh) | — | listed on [skills.sh](https://skills.sh) |

Anthropic’s repo is explicit that GitHub is the demonstration set, not a private DB:

> This repository contains skills that demonstrate what's possible with Claude's skills system. These skills range from creative applications (art, music, design) to technical tasks (testing web apps, MCP server generation) to enterprise workflows (communications, branding, etc.).
>
> — Anthropic, `anthropics/skills` README. [https://github.com/anthropics/skills](https://github.com/anthropics/skills)

Claude Code can register that repo as a plugin marketplace: `/plugin marketplace add anthropics/skills` (same README).

In-repo instruction files are a larger silent corpus than any store. [AGENTS.md](https://agents.md/) is the cross-tool convention:

> AGENTS.md complements this by containing the extra, sometimes detailed context coding agents need: build steps, tests, and conventions that might clutter a README or aren't relevant to human contributors.
>
> — agents.md, retrieved 21 August 2026. [https://agents.md/](https://agents.md/)

The same page notes that nested files win by proximity, and that OpenAI’s own monorepo had on the order of 88 `AGENTS.md` files at the time of writing. Claude Code still prefers `CLAUDE.md`; Copilot still reads `.github/copilot-instructions.md`. Public documentation sites additionally publish `llms.txt` as a machine README (a separate, web-crawler convention — not a skill registry).

## skills.sh is the package manager

Vercel launched the CLI and directory on 20 January 2026:

> We released `skills`, a CLI for installing and managing skill packages for agents. … Today we’re also introducing skills.sh, a directory and leaderboard for skill packages.
>
> — Vercel Changelog, 20 January 2026. [https://vercel.com/changelog/introducing-skills-the-open-agent-skills-ecosystem](https://vercel.com/changelog/introducing-skills-the-open-agent-skills-ecosystem)

Install command (canonical): `npx skills add vercel-labs/agent-skills`

The changelog listed early agent targets: Amp, Antigravity, Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, OpenCode, Windsurf, and others. The CLI README later claims **OpenCode, Claude Code, Codex, Cursor, and 73 more**.

### What the leaderboard actually ranks

On 21 August 2026 the [skills.sh](https://skills.sh) all-time board was not a list of indie prompt packs. The largest install totals were **vendor skill packs rolled into captive seats**:

| Pack / skill cluster (as shown on skills.sh) | Reported installs | Notes |
| --- | --- | --- |
| Feishu / Lark (`open.feishu.cn`, `larksuite/cli`) | ~14.4M + additional Lark clusters at 3.3M–4.5M | Org-wide agent skills for docs, VC, tasks, slides |
| `microsoft/azure-skills` | ~8.0M | Cloud-ops skills (RBAC, AKS, cost, migrate) |
| `prime-skills/runcomfy-agent-skills` | several clusters, 1.7M–3.6M | Video / gen-media |
| `lllllllama/rigorpilot-skills` | ~1.8M + 1.2M | Research / paper context |
| `mattpocock/skills` | ~1.4M | TypeScript teaching skills |
| `obra/superpowers` | ~683k | Agent workflow meta-skills |

Treat these as **distribution numbers, not unique humans**. An enterprise seat that auto-installs a pack counts.

Design skills — the other guess in this survey — are real and they live as GitHub `SKILL.md` folders, not as a Figma store. Vercel ships interface rules as a skill:

> Use these guidelines with AI coding agents. Audit all generated interfaces. … `npx skills add https://github.com/vercel-labs/agent-skills --skill web-design-guidelines`
>
> — Vercel, *Web Interface Guidelines*. [https://vercel.com/design/guidelines](https://vercel.com/design/guidelines)

Related high-visibility design skills on the same leaderboard: `frontend-design`, `anti-ui-slop`, `design-taste-frontend`, `impeccable`, `high-end-visual-design`. Anthropic’s example skills cover “creative applications (art, music, design)” in the public repo quoted above.

## OpenAI does not have one “agent database”

It has a sequence of storefronts. Two of them are winding down.

### Agent Builder

> OpenAI is deprecating Agent Builder. Existing users can continue using it during the transition window, and the product is scheduled to shut down on November 30, 2026. ChatKit remains available.
>
> — OpenAI, *Agent Builder* API docs, retrieved 21 August 2026. [https://developers.openai.com/api/docs/guides/agent-builder](https://developers.openai.com/api/docs/guides/agent-builder)

The company blog update of 3 June 2026 points remaining work at the **Agents SDK** (code) or **Workspace Agents in ChatGPT** (natural language, teams) ([Introducing AgentKit](https://openai.com/index/introducing-agentkit/), update of 3 June 2026).

### Custom GPTs / GPT Store

Press reports dated 16–21 August 2026, citing updated OpenAI Help Center language, state that **personal ChatGPT accounts (Free, Go, Plus, Pro) can no longer create or publish new custom GPTs** as of 16 August 2026; existing GPTs remain usable; Business / Enterprise / Edu workspaces keep creation. Independent confirmation from OpenAI’s HTML help article timed out during this survey; treat the restriction as reported, and verify at [Creating and editing GPTs](https://help.openai.com/en/articles/8554397-creating-a-custom-gpt) before relying on it operationally.

### What OpenAI actually ships now

1. **Skills** — same `SKILL.md` contract. Codex scans `.agents/skills` from the working directory up to the repo root. API: upload via `POST /v1/skills`, attach to the shell/container ([Skills in OpenAI API](https://developers.openai.com/cookbook/examples/skills_in_api); [Build skills](https://developers.openai.com/codex/skills)).
2. **Plugins** — the distribution unit. Skills + MCP connectors + optional UI. Manifest: `.codex-plugin/plugin.json`. Installed from a **universal plugin directory** shared by ChatGPT Work and Codex (`/plugins` in CLI). OpenAI-curated connectors include GitHub, Gmail, Drive, Slack ([Plugins](https://developers.openai.com/codex/plugins)).
3. **Workspace Agents** — team-scoped successors to many GPT use-cases, inside ChatGPT Business / Enterprise / Edu.
4. **Public GitHub catalog** — [openai/skills](https://github.com/openai/skills) (25,084 stars on 21 August 2026).

OpenAI’s own docs say skills are the authoring format and plugins are how you distribute them:

> Skills are the authoring format for reusable workflows. Plugins distribute reusable skills and connectors to ChatGPT Work on the web and to ChatGPT Work and Codex in the desktop app.
>
> — OpenAI, *Build skills* (Codex). [https://developers.openai.com/codex/skills](https://developers.openai.com/codex/skills)

That is not a public skill graph. It is GitHub folders plus a closed directory plus private workspace agents.

## MCP registries: tools, not skills

The official registry is a **metadata feed**, not a code host:

> The MCP Registry is the official centralized metadata repository for publicly accessible MCP servers, backed by major trusted contributors to the MCP ecosystem such as Anthropic, GitHub, PulseMCP, and Microsoft.
>
> The MCP Registry is not intended to be directly consumed by host applications. Instead, host applications should consume other MCP registries, such as downstream marketplaces, via a REST API conforming to the official MCP Registry's OpenAPI spec.
>
> — Model Context Protocol, *About the MCP Registry* (preview). [https://modelcontextprotocol.io/registry/about.md](https://modelcontextprotocol.io/registry/about.md)

Code still lives on npm, PyPI, Docker Hub, and GitHub. Aggregators add search, hosting, and scoring.

| Surface | What it is | Snapshot (21 Aug 2026 unless noted) | URL |
| --- | --- | --- | --- |
| Official MCP Registry | Canonical metadata API; preview; no binaries | preview / API freeze v0.1 announced 24 Oct 2025 | [registry.modelcontextprotocol.io](https://registry.modelcontextprotocol.io) |
| PulseMCP | Human directory; ingest from official registry | **22,029** servers | [pulsemcp.com/servers](https://www.pulsemcp.com/servers) |
| Glama | Broad index + hosted connectors | **75,841** servers, 13,960 connectors, 602,559 tools (Glama’s own counters, “last indexed Aug 21, 2026”) | [glama.ai](https://glama.ai/) |
| Smithery | Marketplace + hosted runtime / OAuth | often cited ~7k hosted; treat as a runtime, not a census | [smithery.ai](https://smithery.ai) |
| Composio | SaaS toolkits behind one MCP gateway | ~1,000 toolkits / ~20,000 tools (vendor + secondary catalogs) | [composio.dev](https://composio.dev) |
| Hugging Face MCP | Hub search + Gradio Spaces as tools | official server | [huggingface.co/docs/hub/agents-mcp](https://huggingface.co/docs/hub/agents-mcp) |
| Context7 | Versioned library docs into the prompt | MCP + skill; commonly cited ~60k GitHub stars | [context7.com](https://context7.com) / [github.com/upstash/context7](https://github.com/upstash/context7) |
| LangSmith MCP | Prompts, traces, datasets from a workspace | hosted `api.smith.langchain.com/mcp` | [LangSmith MCP docs](https://docs.langchain.com/langsmith/langsmith-mcp-server) |
| Docker MCP Catalog | Containerized, reviewed servers | enterprise distribution path | Docker Hub MCP catalogs |
| Cursor Marketplace | Reviewed plugins (skills + MCP + rules) | official | [cursor.com/marketplace](https://cursor.com/marketplace) |
| cursor.directory | Community MCP / plugin gallery | unofficial | [cursor.directory](https://cursor.directory) |

Glama’s 75k and PulseMCP’s 22k are **not the same population**. Aggregators duplicate, include abandoned repos, and count forks. A server in both the official registry and PulseMCP has cleared two listing filters; that overlap is more useful than either headline number.

Context7 is the dominant “library documentation as context” product. Its README:

> Context7 pulls up-to-date, version-specific documentation and code examples straight from the source — and places them directly into your prompt.
>
> — Upstash / Context7 README. [https://github.com/upstash/context7](https://github.com/upstash/context7)

Hugging Face positions the Hub as a context lake plus skills:

> Hugging Face Agents connect AI agents to the Hub. Using MCP (Model Context Protocol), Skills, or open-source tooling, agents can search models, explore datasets, run Spaces, and use community tools.
>
> — Hugging Face, *Agents*. [https://huggingface.co/docs/hub/main/en/agents](https://huggingface.co/docs/hub/main/en/agents)

Google’s Gemini **Managed Agents** API (beta) mounts GitHub sources, `SKILL.md`, and remote MCP into a sandbox ([Building Managed Agents](https://ai.google.dev/gemini-api/docs/custom-agents)). Consumer Gems remain the persona layer.

## Closed persona stores

These still hold a lot of *user-facing* pre-generated context. They are not Agent Skills.

| Store | Shape | 2026 status |
| --- | --- | --- |
| ChatGPT GPT Store | Instructions + knowledge files + actions | Large installed base; **new consumer creation reported blocked** 16 Aug 2026 |
| ChatGPT Workspace Agents | Team agents with tools and skills | OpenAI’s stated NL successor to Agent Builder |
| Gemini Gems / Managed Agents | Instructions + files; API agents mount Git | Gems for people; Managed Agents for developers |
| Poe (Quora) | Multi-model bots, usage-based creator pay | Still a bot marketplace, not a `SKILL.md` index |
| Character.AI | Character + Lorebook / Facts / Story Memory | High volume for character context; Google IP license 2024 |
| PromptBase, FlowGPT, AIPRM | Prompt paragraphs | Lower status once agents want folders and tools |

Cursor’s official storefront is [cursor.com/marketplace](https://cursor.com/marketplace): plugins are git repos with `.cursor-plugin/plugin.json`, manually reviewed. Community MCP still concentrates on [cursor.directory](https://cursor.directory). GitHub Copilot ships custom agents, org-level instructions (GA April 2026 per practitioner writeups), and an in-editor MCP gallery.

## Ranked: where the mass is

If you mean **reusable procedural context that coding agents load**:

1. GitHub (source of truth)
2. [skills.sh](https://skills.sh) (discovery and install counts)
3. Vendor plugin directories — Claude plugins, OpenAI `/plugins`, Cursor Marketplace, Copilot
4. Awesome lists, especially [ComposioHQ/awesome-claude-skills](https://github.com/ComposioHQ/awesome-claude-skills)

If you mean **live or external context the model can pull**:

1. MCP aggregators (PulseMCP, Glama, Smithery)
2. SaaS gateways (Composio; first-party Slack / GitHub / Drive connectors)
3. Docs indexes (Context7, `llms.txt`, Hugging Face Hub)
4. LangSmith and other prompt hubs (mostly team-private)

If you mean **“be this person / know this domain” bots**:

1. GPT Store (legacy; creation gated for consumers)
2. Gems and Workspace Agents
3. Character.AI and Poe

Azure and Lark inflate skill-install charts because they ship packs into every seat. That is not a public commons.

## What is missing

Everything above is **task-shaped**, **repo-shaped**, or **account-shaped**:

- “When I review a PR, do this”
- “In this codebase, run tests like this”
- “You are a travel agent with these PDFs attached”

Almost nobody hosts **place-shaped** context as a first-class object: a fence, a voice, visibility, attestation. The usual hacks are a Custom GPT with a knowledge file named after a city, an MCP that hits a places API, a skill that says “search the web,” or RAG over a wiki. None of those mean: *you are standing here, this is what an insider would say, and it only unlocks on-site.*

That is the hole [Open Context Map](https://opencontextmap.com) is in. The AI world standardized on `SKILL.md` + MCP + `AGENTS.md`. Place-bound, attested context never got a registry.

Practical on-ramps into the existing world, if you are publishing place context:

1. An **MCP server** listed on the official registry so ChatGPT, Claude, and Cursor can discover the *tools*.
2. A **skill** on GitHub + skills.sh that tells agents *when* to call that server (for example: always fetch here-stack first).
3. Not the GPT Store as a growth plan for individuals.

## How to cite

**APA-style**

Open Context Map. (2026, August 21). *Where the AI world hosts pre-generated context*. https://opencontextmap.com/research/2026-08-where-ai-hosts-skills

**BibTeX**

```
@techreport{ocm2026skills,
  author      = {{Open Context Map}},
  title       = {Where the AI World Hosts Pre-Generated Context},
  year        = {2026},
  month       = aug,
  day         = {21},
  institution = {Open Context Map},
  url         = {https://opencontextmap.com/research/2026-08-where-ai-hosts-skills},
  note        = {Survey snapshot. Machine-readable copy: https://opencontextmap.com/research/2026-08-where-ai-hosts-skills.md}
}
```

**For agents quoting a figure.** Name the primary source in the same sentence (for example “PulseMCP listed 22,029 servers on 21 August 2026”), then this article as the survey that collected it. Do not cite this page as the origin of Anthropic’s or OpenAI’s product copy; cite their docs, quoted above.

## Sources

Primary documents (retrieved 21 August 2026 unless noted):

1. Anthropic. *Agent Skills* overview. https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview
2. Anthropic. `anthropics/skills` (GitHub). https://github.com/anthropics/skills
3. Agent Skills specification. https://agentskills.io/specification
4. Vercel. *Introducing skills, the open agent skills ecosystem* (20 Jan 2026). https://vercel.com/changelog/introducing-skills-the-open-agent-skills-ecosystem
5. Vercel Labs. `vercel-labs/agent-skills` and `vercel-labs/skills` (GitHub). Star counts 21 Aug 2026.
6. Vercel. *Web Interface Guidelines*. https://vercel.com/design/guidelines
7. skills.sh leaderboard. https://skills.sh (install totals as displayed 21 Aug 2026)
8. OpenAI. *Agent Builder*. https://developers.openai.com/api/docs/guides/agent-builder
9. OpenAI. *Introducing AgentKit* (update 3 Jun 2026). https://openai.com/index/introducing-agentkit/
10. OpenAI. *Build skills* (Codex). https://developers.openai.com/codex/skills
11. OpenAI. *Plugins*. https://developers.openai.com/codex/plugins
12. OpenAI. `openai/skills` (GitHub). 25,084 stars on 21 Aug 2026. https://github.com/openai/skills
13. OpenAI. *Skills in OpenAI API* (cookbook). https://developers.openai.com/cookbook/examples/skills_in_api
14. Model Context Protocol. *About the MCP Registry*. https://modelcontextprotocol.io/registry/about.md
15. MCP Registry (GitHub). https://github.com/modelcontextprotocol/registry
16. PulseMCP server directory. 22,029 servers on 21 Aug 2026. https://www.pulsemcp.com/servers
17. Glama MCP index. 75,841 servers, last indexed 21 Aug 2026 (site counters). https://glama.ai/
18. ComposioHQ. `awesome-claude-skills` README. 67,459 stars. https://github.com/ComposioHQ/awesome-claude-skills
19. Hugging Face. *Agents* and *MCP Server*. https://huggingface.co/docs/hub/main/en/agents and https://huggingface.co/docs/hub/agents-mcp
20. Hugging Face. `huggingface/skills`. https://github.com/huggingface/skills
21. Upstash. Context7 README. https://github.com/upstash/context7
22. LangSmith MCP server docs. https://docs.langchain.com/langsmith/langsmith-mcp-server
23. AGENTS.md. https://agents.md/
24. Cursor. Skills help. https://cursor.com/docs/skills and https://cursor.com/help/customization/skills
25. Cursor Marketplace. https://cursor.com/marketplace
26. Google. *Building Managed Agents* (Gemini API). https://ai.google.dev/gemini-api/docs/custom-agents
27. OpenAI Help Center. *Creating and editing GPTs*. https://help.openai.com/en/articles/8554397-creating-and-editing-gpts — fetch timed out during this survey; consumer GPT-creation restriction cited from contemporaneous press (16–21 Aug 2026) pending re-check of that page.

Secondary catalogs used only for cross-checks: TrueFoundry *Best MCP Registries in 2026*; Firecrawl *Best Codex Plugins* / *Best MCP Servers for Gemini CLI*; DesignRevision MCP marketplaces roundup.

---

*Open Context Map is a public atlas of place-bound AI context. This note is a survey, not an endorsement of any marketplace. Corrections: the Research page on this site.*