The Future of AI Agent Plugins: From GPT Store to MCP

The idea of extending AI agents with third-party capabilities has been tried multiple times. Understanding the history helps explain why the current approach — MCP for tool access, SKILL.md for agent behavior — is sticking where earlier attempts didn't.

The ChatGPT plugins era (2023)

OpenAI launched ChatGPT plugins in March 2023 with significant fanfare. The concept was simple: third-party developers could build plugins that gave ChatGPT access to external services. Expedia for travel, Wolfram Alpha for math, Zapier for automations.

Plugins didn't take off for several reasons. They were expensive to build and maintain. Distribution was controlled entirely by OpenAI. The user experience was clunky — you had to manually enable plugins for each conversation. And most importantly, the plugins couldn't do much that a simple web search couldn't handle.

By late 2023, OpenAI quietly deprecated plugins.

The GPT Store era (2024)

The GPT Store replaced plugins with custom GPTs — simplified chatbots with custom instructions, knowledge files, and limited tool access. The store launched in January 2024 and attracted millions of creators.

The GPT Store solved distribution better than plugins, but the capabilities were limited. Custom GPTs couldn't access local files, execute code on your machine, or integrate with your development workflow. They were useful for specialized conversations but not for real work.

The MCP and SKILL.md era (2025-2026)

Two standards emerged that actually solve the extensibility problem:

MCP (Model Context Protocol) handles what plugins tried to do — connecting agents to external tools and services. But MCP is vendor-neutral, runs locally, and gives developers fine-grained control over what their agent can access. Any agent can connect to any MCP server.

SKILL.md handles what custom GPTs tried to do — teaching agents how to behave for specific tasks. But SKILL.md skills are portable across agents, stored locally, and integrate directly into the development workflow.

Together, MCP and SKILL.md cover the full spectrum of agent extensibility without the walled-garden problems that killed earlier approaches.

Why the open standard approach works

The critical difference is vendor neutrality. ChatGPT plugins only worked with ChatGPT. Custom GPTs only work with ChatGPT. MCP and SKILL.md work with Claude Code, Codex CLI, Cursor, Gemini CLI, VS Code, and every other compatible agent.

This means:

• Creators have a larger addressable market
• Users aren't locked into one agent
• Competition happens on quality, not platform lock-in
• The ecosystem can grow faster because innovation isn't bottlenecked by one company

Where it's heading

The agent plugin ecosystem is converging on MCP for tool access and SKILL.md for behavior. The remaining challenge is discovery and distribution — finding the right MCP servers and skills for your specific needs.

This is where marketplaces like Agensi come in. They aggregate, curate, and distribute agent capabilities across the ecosystem, providing the discovery layer that GitHub and npm don't optimize for.

The future likely includes more specialized skills (industry-specific, framework-specific), better quality signals (ratings, verified installs, security audits), and enterprise features (private registries, access controls, compliance). But the foundation — open standards for tool access and agent behavior — is set.