When ChatGPT launched plugins in 2023, it felt like a big deal — and it was, for about a year. Then OpenAI deprecated them. Meanwhile, Anthropic built something that works differently at a fundamental level, and the comparison keeps coming up. So here's the honest breakdown of what MCP servers and plugins actually are, why they're different, and why it matters to you.
What plugins were (and why they went away)
ChatGPT plugins were a system where third-party developers could register their API as a "plugin" that ChatGPT could call during a conversation. To build one, you needed OpenAI's approval. To use one, you needed to be on ChatGPT Plus. And if OpenAI decided to deprecate the system — which they did — your plugin stopped working.
Plugins ran on the internet. They called external APIs. They couldn't touch your local machine. And every call you made went through OpenAI's servers before reaching the plugin's server. That's three layers of infrastructure between you and your tool.
They were useful but fragile. The marketplace grew to thousands of plugins, most of which were low quality. OpenAI replaced them with GPTs in 2024.
What MCP servers actually are — and how they're different
MCP is an open protocol, not a vendor marketplace. Anyone can implement it. No approval process. No marketplace listing required. You build a server that speaks the MCP protocol, and any MCP-compatible client — Claude Desktop, Cursor, Zed, or anything else — can use it.
And crucially: MCP servers run where you choose. You can run them on your own machine, in your own cloud account, or in a trusted hosted environment. There's no mandatory routing through anyone else's servers.
Side-by-side comparison
| Feature | AI Plugins (e.g., ChatGPT) | MCP Servers |
|---|---|---|
| Open standard? | No — proprietary per vendor | Yes — open spec, Apache 2.0 |
| Can run locally? | No — always internet-hosted | Yes — runs on your machine |
| Requires vendor approval? | Yes | No |
| Works across AI clients? | No — single platform | Yes — Claude, Cursor, Zed, etc. |
| Local file access? | No | Yes (filesystem server) |
| Local database access? | No | Yes (Postgres, SQLite) |
| User controls data routing? | No | Yes |
| Startup effort? | One click | Config file edit + restart |
| Discovery? | Curated marketplace | GitHub, community directories |
| Security model? | Vendor-defined | User-controlled |
Where plugins beat MCP servers (honestly)
Plugins were easier to install — literally one click from a curated store. MCP servers require editing a JSON file and restarting an app. That's a real barrier for non-technical users.
Plugin marketplaces also made discovery trivially easy. With MCP, you need to know GitHub, understand package names, and read READMEs. The community is working on making this better, but it's not there yet.
So if you're comparing for sheer ease of use, plugins win. But if you care about what you can actually do — especially local access to files, databases, and code — MCP is in a different league.
MCP servers vs. Claude Skills — another comparison worth making
Here's a distinction many people miss: MCP servers are different from Claude Skills too. Claude Skills are pre-built, Anthropic-curated capabilities (like "run Python code" or "create a spreadsheet") that you toggle on in the Claude interface. They're convenient but limited to what Anthropic ships.
MCP servers are user-defined and can do anything a developer builds. They're more powerful and more flexible — but they require your own setup. For the full breakdown, see our guide on what Claude Skills are.
Are MCP servers replacing tools/functions in the Claude API?
No. The Claude API's tool_use feature (where you define custom functions in your API request) is still the right approach when you're building an application that calls Claude programmatically. MCP is primarily for interactive use in clients like Claude Desktop, Cursor, and Zed. The two approaches coexist — you'd use MCP for interactive tools and API tool_use for programmatic integrations.
Frequently Asked Questions
No. ChatGPT plugins were a proprietary OpenAI standard that required vendor approval and were tied to the ChatGPT interface. MCP is an open protocol that any developer can implement, working across multiple AI clients — not just one.
Claude didn't have a plugins system in the way ChatGPT did. Anthropic went straight to MCP, which launched in November 2024, as their standard for tool integration.
For most technical use cases, yes — more powerful, more private, more flexible. For non-technical users who want one-click setup without config files, the plugin model was simpler. MCP is catching up on ease of use as tooling improves.
MCP is an open protocol, so theoretically any AI client can implement it. As of 2026, Claude Desktop, Cursor, Zed, and several other tools support MCP. OpenAI and Google have their own tool integration approaches but haven't adopted MCP — yet.