The protocols that turned AI models into an agent economy
In 20 months the agentic stack went from zero to standardized: one protocol for tools, one for agent-to-agent collaboration, one for interfaces, two for payments — and an open format for packaging expertise. This is the deep, visual guide to the state of the art, with live adoption data — plus a hands-on case study of one real working setup.
One stack, six layers
Every acronym in agentic AI fits on this map. The model reasons at the bottom; each layer above gives it a new power — knowledge, hands, colleagues, a face, and finally a wallet. Every layer now has an open protocol.
Tools, APIs, MCP, Skills — what's actually the difference?
The framing top explainers converge on: tools are the hands, MCP is the port, skills are the playbooks, protocols are the language. An agent is a model given agency (it decides), autonomy (it acts without step-by-step supervision), and self-governance (it stays within your rules) — the three properties Biswas & Talukdar use to define agentic systems.
Tools the hands
A tool is one function the model may call — search_flights(date, city), run_sql(query). The model reads the tool’s schema, decides when to call it, and gets structured results back. Tool calling is the atom of agency: without tools, an LLM can only talk.
APIs the raw wiring
APIs are how software talks to software — REST, GraphQL, gRPC. Agents consume them through tools. The problem: every API is different, so every integration was custom glue code. That N×M explosion is exactly what MCP was invented to kill.
MCP the universal port
The Model Context Protocol standardizes how models discover and call tools, read resources, and reuse prompts. Build one MCP server for your database and every MCP client — Claude, ChatGPT, Cursor, your own app — can use it. USB-C for AI.
Skills the playbooks
A skill is a folder with a SKILL.md — instructions, scripts, references — that teaches an agent how to do something well: build slide decks, review contracts, deploy to Vercel. Loaded only when relevant (progressive disclosure), so 500 skills cost almost nothing until one is needed.
Subagents the team
Agents that spawn specialized agents — a planner, a coder, a reviewer — each with its own context and tools. The book calls this the Coordinator–Worker–Delegator pattern: one orchestrator, many focused workers, results merged back.
The agent loop the heartbeat
Every agent runs the same cycle: reason about the goal, act with a tool, observe the result, reflect, repeat (ReAct). Reflection and introspection — the agent critiquing its own output — is what separates an agent from an autocomplete.
Cheat sheet: which piece do you need?
One decision table instead of forty tutorials.
| You want the model to… | Reach for… | Example |
|---|---|---|
| Do one concrete action | a Tool | send_email(to, subject, body) |
| Connect to a whole system, reusably | an MCP server | Supabase MCP: 20+ tools from one connection |
| Follow your procedure & standards | a Skill | slides skill: how YOU build decks |
| Handle a big task in parallel | Subagents | planner + coder + reviewer |
| Talk to someone else’s agent | A2A | your agent hires a travel agent’s agent |
| Show real UI, not text walls | A2UI / AG-UI | date pickers, live progress, approvals |
| Spend money safely | AP2 / x402 | signed mandates; per-call micropayments |
The seven protocols, one at a time
Each layer of the stack got its own open standard between late 2024 and late 2025. Pick a protocol — every panel has the plain-English version, the architecture, live adoption numbers, and links to the primary sources.
MCP — Model Context Protocol
Agent ↔ Tools & DataAnthropic → open ecosystem (OpenAI, Google, Microsoft adopted) · Born Nov 25, 2024
The open standard that connects an AI model to the outside world: files, databases, SaaS apps, browsers — anything. One protocol instead of an N×M mess of custom integrations.
How it works
- Client–server over JSON-RPC 2.0: your AI app (host) runs MCP clients that connect to MCP servers exposing capabilities.
- Three primitives — Tools (actions the model can call), Resources (data it can read), Prompts (reusable templates).
- Two transports: stdio for local servers, Streamable HTTP for remote ones, with OAuth 2.1 for authorization.
- The model discovers what a server offers at runtime — no hardcoding. New server, new powers, zero code changes.
By the numbers
Moving parts
- Host app (Claude, ChatGPT, IDEs)
- MCP client (one per connection)
- MCP server (wraps a tool/data source)
- JSON-RPC 2.0 messages over stdio / HTTP
The wider landscape
Standards worth knowing that orbit the core seven.
Adoption, in real numbers
Popularity is measurable: repository gravity, partner counts, governance moves. Here is where each protocol actually stands — plus what the analysts project.
Ecosystem gravity — GitHub stars per protocol / SDK repo
Fetched live from the GitHub API on July 17, 2026. Linear scale — the disparity is the story.
Read: packaged expertise (skills) and tool connectivity (MCP) dominate developer attention; agent-to-agent and payments layers are earlier on the curve. ACP is dormant — it merged into A2A.
20 months from zero to standard
Every layer of the stack got an open protocol in under two years.
🏆 The Top 20 AI Agent Arsenal (2026)
The most-starred tools, skill packs and harnesses in the global agent ecosystem — every repository verified against the GitHub API on July 17, 2026, and ranked by real star counts, not by the stale numbers circulating in listicles.
| # | Tool / Skill pack | Stars (live) | Primary utility | Repository | |
|---|---|---|---|---|---|
| 1 | OpenClaw | 383,206 ★ | The #1 open-source personal AI assistant — any OS, any platform, the lobster way 🦞. | openclaw/openclaw ↗ | |
| 2 | Superpowers | 256,206 ★ | Agentic skills framework enforcing TDD, brainstorming-first and disciplined planning. | obra/superpowers ↗ | |
| 3 | Everything Claude Code (ECC) | 230,454 ★ | Agent harness optimization: skills, instincts, memory, security, hooks. (Repo renamed from everything-claude-code.) | affaan-m/ECC ↗ | |
| 4 | n8n | 196,757 ★ | Fair-code workflow automation with native AI agent nodes — the no-code gateway to agentic systems. | n8n-io/n8n ↗ | |
| 5 | Ollama | 176,293 ★ | The backbone for running frontier open models locally (Kimi, GLM, DeepSeek, Qwen…). | ollama/ollama ↗ | |
| 6 | Firecrawl | 152,147 ★ | The API to search, scrape and turn any website into LLM-clean Markdown at scale. | firecrawl/firecrawl ↗ | |
| 7 | Dify | 149,118 ★ | Production-ready platform for building and managing agentic workflow backends. | langgenius/dify ↗ | |
| 8 | Open WebUI | 145,713 ★ | The most popular local-first chat interface for LLMs (Ollama, OpenAI-compatible). | open-webui/open-webui ↗ | |
| 9 | UI UX Pro Max | 106,748 ★ | Design-intelligence skill for building UIs that don’t look AI-generated. | nextlevelbuilder/ui-ux-pro-max-skill ↗ | |
| 10 | Scrapling | 69,819 ★ | Adaptive web scraping framework that auto-adjusts when sites change. | D4Vinci/Scrapling ↗ | |
| 11 | Get Shit Done (GSD) | 64,761 ★ | Meta-prompting, context-engineering and spec-driven development framework. | gsd-build/get-shit-done ↗ | |
| 12 | Awesome Claude Code | 50,210 ★ | The master directory of skills, hooks and resources for Claude Code. | hesreallyhim/awesome-claude-code ↗ | |
| 13 | Impeccable | 47,431 ★ | The design language that makes your agent better at design (/audit, /polish). | pbakaus/impeccable ↗ | |
| 14 | Claude Skills (Fullstack) | 10,615 ★ | 66 specialized full-stack skills — framework-specific deep knowledge packs. | Jeffallan/claude-skills ↗ | |
| 15 | Magic MCP | 5,452 ★ | v0-style UI component generation injected straight into your IDE via MCP. | 21st-dev/magic-mcp ↗ | |
| 16 | Obsidian Wiki | 2,891 ★ | Framework for agents to build and maintain a digital brain in Obsidian. | Ar9av/obsidian-wiki ↗ | |
| 17 | TÂCHES CC Resources | 1,973 ★ | Curated custom Claude Code resources; meta-skill auditing against prompt bloat. | glittercowboy/taches-cc-resources ↗ | |
| 18 | Claude Pace | 214 ★ | Statusline rate-limit tracker with pace-aware quota monitoring. Pure Bash. | Astro-Han/claude-pace ↗ | |
| 19 | Stitch Kit | 35 ★ | Design-to-code bridge on Google’s Stitch engine — 35 skills, 7 framework targets. | gabelul/stitch-kit ↗ | |
| 20 | notebooklm-py | 4 ★ | Full Python API access to NotebookLM beyond the official UI. | patriarchal-boothose896/notebooklm-py ↗ |
Corrections vs. lists circulating online: everything-claude-code is now affaan-m/ECC, Firecrawl moved to the firecrawl org, Magic MCP lives at 21st-dev/magic-mcp, and most star counts out there are months stale — OpenClaw alone gained six figures. Always verify against the API.
Use them today: skills & MCP in three moves each
The two pieces you can adopt this afternoon. No framework, no PhD — a folder and a config entry.
📘 Add a skill to Claude Code
Create the folder
A skill is just a directory with a SKILL.md inside.
mkdir -p ~/.claude/skills/deck-builder
Write SKILL.md — frontmatter + instructions
The description decides when the agent loads it. That one line is 80% of the craft.
--- name: deck-builder description: Use when building slide decks. Applies my structure, brand and chart rules. --- # Deck Builder 1. Ask for audience + duration first. 2. One idea per slide, max 12 slides...
Just ask
Say "build me a deck about X" — the agent sees the skill's description, loads the full instructions only then. That's progressive disclosure: hundreds of installed skills cost ~40 tokens each until one activates.
🔌 Connect an MCP server
Pick a server
Thousands exist: GitHub, Supabase, Playwright, Slack, Stripe, Figma… Browse the official registry or the servers repo.
# one command in Claude Code: claude mcp add github \ -- npx -y @modelcontextprotocol/server-github
Authorize it
Remote servers use OAuth 2.1 — you approve scopes once; local servers run on stdio with your credentials in env vars, never in code.
The model discovers the rest
Tools, resources and prompts are self-describing. Ask "list my open PRs" and the model finds search_pull_requests on its own. New server version, new tools — zero changes on your side.
The rule of thumb the tutorials agree on
MCP for connections, skills for conventions. MCP gives the agent access to systems (what it can touch); skills encode how your team wants things done (how it should behave). They compose: a deploy skill that calls the Vercel MCP server is the whole stack in one workflow.
Build a real SKILL.md, right here
A skill is a folder with a SKILL.md file inside. Six parts do all the work — learn what each one is for, then fill in the form and copy a working skill you can install in one minute.
1 · Frontmatter: name required
Lowercase letters, numbers and hyphens only (max 64 chars). It becomes the folder name and how the agent refers to the skill — deck-builder, not "My Deck Builder!".
2 · Frontmatter: description the trigger
The single most important line (max ~1024 chars). It is the ONLY part the agent always sees, and it decides when the skill activates. Formula: what it does + when to use it — "Use when the user asks for a slide deck. Applies my structure, brand and chart rules."
3 · Body: When to Use scope
A short section listing trigger cases and — just as important — when NOT to use the skill. This stops the agent from over-applying it to everything.
4 · Body: Instructions the playbook
Numbered steps written like you would brief a new teammate: exact order, decisions to make, quality bar, what "done" means. Concrete beats clever — "max 12 slides, one idea each" outperforms "make it good".
5 · Body: Examples & Guardrails calibration
One or two input → expected-output examples calibrate the agent better than paragraphs of prose. Guardrails list hard limits: never invent data, always ask audience first, keep files under 800 lines.
6 · Optional: references/ & scripts/ on demand
Big reference docs go in references/*.md and executable helpers in scripts/ — the agent reads them only when needed. This is level 3 of progressive disclosure: your SKILL.md stays lean.
How the agent actually reads it — progressive disclosure
Level 1: at session start the agent sees only name + description (~40 tokens per skill — that's why 500 installed skills cost almost nothing). Level 2: when your request matches the description, the full SKILL.md body loads. Level 3: references/ and scripts/ are read only if the task needs them. Write each level accordingly: razor-sharp description, lean body, deep references.
--- name: deck-builder description: Builds slide decks with my structure, brand colors and chart rules. Use when the user asks for a presentation, slide deck, or pitch. --- # Deck Builder ## When to use - Use when the user asks for a presentation, slide deck, or pitch. - Do NOT use for unrelated requests — stay in scope. ## Instructions 1. Ask for the audience, duration and goal before writing anything 2. Outline max 12 slides — one idea per slide 3. Write the deck as a single HTML file with speaker notes 4. End with a summary slide and one clear call to action ## Example **User:** "Build me a 10-minute deck about AI agents for HR directors" **Agent:** applies every step above, in order, before delivering. ## Guardrails - Never invent statistics — ask for sources or mark as placeholder - Keep text per slide under 40 words
Install it (Claude Code)
mkdir -p ~/.claude/skills/deck-builder # paste the copied content into: # ~/.claude/skills/deck-builder/SKILL.md # then just ask: "Build me a 10-minute deck about AI agents for HR…"
The library: papers, specs & guides
Everything on this site traces back to a primary source. Start with the two surveys, then go protocol by protocol.
📄 Papers & books
📚 Specs & engineering guides
Case study: one practitioner's skill library, measured
Everything above is the global state of the art. This section is different: a concrete example of what one real working setup looks like — a live scan of the skills installed on the author's machine (1,654 SKILL.md files, deduplicated to 523 unique skills, including 13 of the Top 20 arsenal). Use it to calibrate what "adopting the stack" means in practice, not as an industry reference.
Top 10 heaviest skills by token weight
Estimated as SKILL.md bytes ÷ 4 · scanned 2026-07-17
Most used in practice
Curated from real workflows: teaching decks, research, shipping Vercel apps.
slidesHTML slide decks for teaching and client workshopsteachingdeep-researchMulti-source cited research via Firecrawl + Exaresearchmarket-researchMarket sizing & competitive analysis for AplicAppresearchfrontend-designDistinctive UI for every app shippedbuildinggraphifyTurns any input into knowledge graphs (his largest skill: ~14.4k tokens)buildingcopywritingLanding copy that convertsmarketingprompt-engineerSharper prompts for agent pipelinesagentscreate-agent-skillsMeta: skills that write skillsagentsvercel:deployShip to production from the terminalbuildingtoken-budget-advisorCustom-built: controls response depth & costagents