Comparisons · How agents get their tools
ChatGPT started the plugin era. Even OpenAI moved to MCP.
For a year and a half, giving an AI a new skill meant building it the way ChatGPT did in 2023: a plugin, wired to one vendor's cloud, usable only inside that one app. Then Anthropic published the Model Context Protocol (an open standard any agent can speak) and the math inverted. Build a capability once, and every MCP agent can use it: Claude Code, Cursor, and by 2025, ChatGPT itself. Here is the difference between the two models, drawn out, and why a single install can hand the same file search to 19 different AI clients.
The short version
- 01ChatGPT plugins (2023) let one app (ChatGPT) call outside tools. Every other assistant needed its own separate, incompatible format.
- 02MCP flips it: one server exposes a capability, and every MCP agent (Claude Code, Cursor, ChatGPT and more) can use it. Build once, reach all.
- 03By 2025, even OpenAI, which started the plugin era, had added MCP support to its own products. The neutral standard won.
- 04Plugins tie a tool to a vendor; MCP ties it to a protocol. Switch editors and an MCP tool comes with you; a plugin gets left behind.
- 05Interlinked Files is one MCP server: a single install reaches 19 AI clients and answers file search in microseconds, 85 µs across 4.47M files.
Definitions first
What each one actually is.
ChatGPT plugins were OpenAI's 2023 system for giving ChatGPT access to outside tools: third-party services the model could call mid-conversation, defined to OpenAI's own spec and usable only inside ChatGPT.
MCP is an open standard that lets any AI agent connect to any tool or data source through one shared interface: write the connection once, and every agent that speaks MCP can use it, no per-vendor rebuild.
- Mar 2023OpenAI launches ChatGPT plugins in beta. Tools live inside ChatGPT, and nowhere else.
- Late 2023 to 2024OpenAI shifts to GPTs and the GPT Store, and winds the original plugins down.
- Nov 2024Anthropic introduces the Model Context Protocol as an open standard, and open-sources it.
- 2025Claude Code, Cursor and a wave of agents ship native MCP; OpenAI adds MCP to its own products.
- TodayOne MCP server reaches 19 AI clients in a single install. New agents that speak MCP just work.
The clearest signal is the last step: the company that popularized AI plugins now speaks MCP too. When the incumbent adopts the neutral standard, the debate is over, not because plugins were bad, but because a tool you build for whatever agent you happen to use beats a tool locked to one. For the full primer, see what is an MCP server.
The two models, drawn
Three vendors, three formats. Or one server, every agent.
On the left, the plugin model: each vendor invents its own format, and each capability is trapped inside the one app that hosts it. Want “find files” in a second assistant? Build it again, its way. On the right, MCP moves the capability outside every app into one small server; each agent (including ChatGPT, which adopted MCP in 2025) connects to the same one.
The left side is why tool makers used to pick a favorite app and ignore the rest: every new host meant a fresh integration to write and maintain. The right side is why that math changed: the server is written against the protocol, not against any one app, so the agent on the other end is interchangeable. This is the same shift covered in the backbone of the LLM economy.
Side by side
Six questions, two very different answers.
| The question | ChatGPT-style plugin | MCP server |
|---|---|---|
| Which apps can use it | Only its host, e.g. ChatGPT itself | Every MCP agent on the machine |
| Who defines the format | Each vendor, privately | One open standard, shared |
| To support a new agent | Build a whole new plugin | It already works |
| You switch agents | Left behind | Comes with you |
| Where it runs | The vendor's cloud, behind the app | Its own process, on your machine |
| Do your files leave the machine | Often: the tool is vendor-hosted | No: local server, offline by default |
None of this makes plugins useless: a deep, app-specific extension is still the right tool when you want to reshape one editor. But when the job is handing a capability to whatever agent you happen to be using, the plugin model asks you to bet on which vendor wins. MCP lets you skip the bet, and keeps your files local while it does.
Measured, not claimed
What one shared server buys.
| The searcher | What ran | Time | vs Interlinked |
|---|---|---|---|
| Interlinked Files | one indexed filename lookup | 85 µs | baseline |
| VS Code (Ctrl+P) | same file, same drive | 35 s | 412,000× |
| Windows Search | median of 20 queries | 67 s | ~480,000× |
| ripgrep | same content query, same corpus | 93.8 s | 58,625× |
| Claude Code (no index) | one file, agentic archaeology | 6 m 57 s | ~7,200,000× |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Each ratio is against one indexed Interlinked lookup on the same corpus: filename lookups are microsecond-scale (85 µs single, 139 µs median of 20 queries); the ripgrep row is content-to-content (93.8 s → 16 ms; 58,625× across the query set); the Claude Code row is a measured single-file agent run with no index, listing directories until it found the file. Every one of those numbers is identical no matter which of the 19 clients made the call.
The file-search MCP server
Interlinked Files is one server every agent shares.
It exposes four tools (filename search, content search, both at once, and a knowledge lookup that returns everything known about one file, folder, or repo) over the same protocol Claude Code, Cursor and ChatGPT already speak. There is no ChatGPT plugin to publish, no per-vendor format, no account for the search itself: one signed installer configures 19 AI clients in a single step, and the moment a new MCP agent ships, it works too. Backed by an always-fresh index of every file on your PC (every repo, every document, even the git-ignored files a vendor-hosted plugin never sees), measured on a real 4.47-million-file machine. It stays on your machine, offline, sitting around 44 MB at rest (less than a single browser tab), with the index taking under 1% of the drive. Save a file and it is searchable again in about a millisecond.
ONE SERVER TO INSTALL. NINETEEN AGENTS THAT GET IT. ZERO PER-VENDOR PLUGINS.
Common questions
MCP vs ChatGPT plugins, answered.
Ready to wire it up? The step-by-step is in how to set up an MCP server in Claude Code.
Do this today
Skip the plugins. Install the layer.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No plugin to publish, no per-vendor JSON to edit.
Let it index. Filename search works immediately; content search fills in behind it. At rest the whole engine sits around 44 MB (less than a browser tab), and stays entirely on your machine.
Ask from any agent. Whichever agent you happen to be in, the tools are already there and answer identically. Switch editors tomorrow and there is nothing to reinstall.
It stays fresh. Save a file and it's searchable in about a millisecond (under 30 in the worst case). The index rides your changes; you never re-index by hand.
Everything local is free forever, no card. Hosting starts at $5.99/mo. Download for Windows.