OpenAI launched ChatGPT Plugins in March 2023 and deprecated them thirteen months later. The plugin model — a static manifest, a handshake API, a user clicking “install” — died before it ever reached escape velocity. Its replacement, already reshaping the competitive landscape, is the agent skill: a composable, task-specific capability the AI discovers, selects, and chains at runtime without a human in the loop. This layer — not the base model — will determine which platforms win.
Plugin architectures treated the LLM as a consumer of predefined integrations. Skill architectures treat the LLM as an orchestrator that reasons about available capabilities and composes them on the fly. That distinction has shifted the competitive dynamics of the entire AI stack. The model frontier is flattening. The skill ecosystem is not.
The Plugin Dream Died in 13 Months
ChatGPT Plugins were well designed for their era. Developers wrote a manifest, exposed an API endpoint, and users discovered tools in a storefront. By April 2024, OpenAI had moved on to GPTs — a simpler, more controlled capability model. The reason was architectural. Plugins made the user the router — discovering, installing, and activating tools manually. The model never reasoned about tools; it consumed their output when triggered.
Compare that with Anthropic’s tool use GA announcement on May 30, 2024. StudyFetch CTO Ryan Trattner reported a “42% increase in positive human feedback” after implementing tool use in their AI tutor Spark.E.
The critical difference: Spark.E discovers tools in context, selects the right one, and chains multiple capabilities into a single interaction. The user doesn’t click “install plugin.” They ask a question. — Ryan Trattner, StudyFetch CTO
Those are fundamentally different bets about where intelligence lives in the system.
Why Skills Are Different: Composition Beats Installation
Skills don’t just replace plugin installation with autonomous selection — they enable composition. The model chains multiple skills together in a single reasoning pass. A coding agent invokes a filesystem server, a search tool, and a code formatter in sequence without the user touching any of them.
The numbers show which approach the ecosystem chose:
- The Model Context Protocol (MCP), announced by Anthropic on November 25, 2024, has grown into the dominant standard for agent-to-tool connectivity
- MCP servers GitHub repo: 86.7K stars
- Python SDK: 23.2K stars
- TypeScript SDK: 12.6K stars
- MCP organization: 47.6K followers
- Skills Over MCP working group
- LangChain : 138K stars
- n8n : 191K stars, 400+ integrations, 9,500+ templates
Every one of these projects independently converged on the same abstraction: a skill or tool the agent can discover, reason about, and invoke at runtime. None of them look like the plugin model.
The App Store for Agents Already Exists — It’s Just Distributed
The dominant narrative about AI platform lock-in centers on the model. “Claude is better at coding.” “GPT-5 is better at reasoning.” This framing grows less relevant by the month. As models converge on capability, the switching cost is not the model — it’s the skill ecosystem built on top of it.
Anthropic maintains a dedicated Skills product page . The Claude Skills Collection catalogs 337 Claude Code skills, 30+ agents, and 70+ custom commands. OpenCode — 160K+ GitHub stars — lets developers define custom subagents as markdown files. Antigravity , Google’s agent-first platform, ships with 1,400+ agentic skills.
The plugin model had a centralized store. The skill model has Git. Distribution is embedded in the development workflow itself.
“Open technologies like the Model Context Protocol are the bridges that connect AI to real-world applications, ensuring innovation is accessible, transparent, and rooted in collaboration.” — Dhanji R. Prasanna, CTO at Block
Three Bets Every Builder Must Place Now
- Your plugin integration strategy is obsolescent. If you are building a “plugin” for an AI platform, you are building for an architecture the industry has already rejected. Build an MCP server instead. It works across Claude, OpenCode, Cursor, and any MCP-compatible client.
- The skill ecosystem is your real vendor lock-in — not the model. Switching from Claude to GPT-5 with 30 MCP servers wired into your CI/CD is a config change. Switching from MCP to a proprietary skill protocol is a rewrite.
- Skill discovery is 2026’s app store problem. The MCP Registry (6.9K stars) and Composio (1,000+ toolkits) are early approaches. The winners will make skill discovery as reliable as package managers made library discovery.
The plugin era was a rehearsal. The skill era is the performance. The composable capability layer on top of commoditized models is where competitive advantage lives. Bet accordingly.
Further Reading
- Building Production-Ready Claude Code Skills — Practical guide to developing, testing, and distributing Claude Code skills at scale
- Understanding MCP: The Model Context Protocol Explained — Deep dive into how MCP enables agent-to-tool connectivity
- The Abstraction Tax: Measuring What MCP Costs Against CLI and Skills — Performance analysis of MCP vs direct CLI invocation
- Anthropic Acquires Stainless: The MCP Infrastructure Play — What the acquisition means for the MCP ecosystem
- Coding Agents in 2026: Codex vs Claude Code vs Antigravity vs Copilot — Head-to-head comparison of major AI coding agent platforms
No comments yet