VMGM Software
Start here
Blog7 min

What are Claude connectors and ChatGPT plugins?

Claude calls it a connector, ChatGPT calls it a plugin, and developers call it an MCP server. It is one thing, you build it once, and here is what it takes to get listed in both.

Portrait of Gary Miller

Gary Miller · Co-founder, VMGM Software

Your customers already have ChatGPT or Claude open all day. Sooner or later one of them asks why they can't just tell it to do the thing in your product. When you go to find out how, you run into Claude connectors, ChatGPT plugins and MCP servers, and it turns out they are one thing.

One thing, several names

  • MCP server, to developers: a server that speaks the Model Context Protocol, the open standard Claude, ChatGPT and Cursor all use.
  • Connector, in Claude: an MCP server added to Claude, from its Connectors Directory or by pasting in a URL.
  • Plugin, in Claude: a bundle of one or more MCP servers plus skills, the written instructions that teach Claude how to use them.
  • Plugin, in ChatGPT: an MCP server, called a connector until December 2025, an app after that, and listed in a Plugin Directory since July.

ChatGPT also had a feature called plugins in 2023. It was retired, and the new Plugin Directory is a different thing.

Underneath every one of these names is an MCP server, and one MCP server can be listed in Claude, in ChatGPT, and in any other assistant that supports the protocol, including Cursor and Codex.

What an MCP server is

An MCP server is a menu your product hands to an AI. Each item on it is a tool: a name, a sentence saying what it does, and the inputs it needs. The assistant reads the menu, decides which tool fits what the person asked for, and calls it.

Take View Printer, our social media scheduler. Its MCP server has tools to list connected accounts, upload media, schedule a post and check what went out. Someone types "schedule these three posts across next week at 9am Chicago time," and the assistant schedules them with those tools. They never open View Printer.

Sign-in works the way "Sign in with Google" does. The first time someone uses your product from ChatGPT or Claude, they click connect, sign in to your product in their browser, and approve access. From then on the assistant acts as that person, with that person's permissions, and nothing more. The standard for this is OAuth, and both directories require it for any product that has user accounts.

What it takes to build one

An API, or something close to one. The server calls your product the same way your own app does. If your product has a public or internal API, most of the work is already done.

A server on the internet. Both directories only list remote servers, reached over HTTPS. It runs in your cloud next to the product it talks to.

Tools designed for the jobs, not copied from the API. An API with 80 endpoints should not become 80 tools. The assistant chooses better from a dozen tools shaped around what people ask for, each with a description written for a model to read.

Every tool labeled read or write. Both directories require each tool to declare whether it only reads or whether it changes something, and whether the change can be undone. The assistant uses those labels to decide when to ask the person before it acts.

A test account. Reviewers at Anthropic and OpenAI sign in and try every tool. They need an account with real-looking data in it.

Getting listed in the Claude Connectors Directory

You submit through Anthropic's developer portal at claude.ai/directory/manage, from any paid Claude plan. It asks for:

  • The server's https:// URL. It reads your tools from the server and flags any without a title or a read/write label.
  • A listing: name, one-line summary, description, categories, icon.
  • Documentation, privacy policy and support contact.
  • How users sign in, and whether the data is yours or a partner's.
  • Test-account credentials and instructions for the reviewers.
  • Confirmation that you ran every tool yourself in Claude first.

Anthropic scans each submission automatically and lists it as a Community connector by default. Some submissions also get a review from a person, and listings can later earn a Verified label. Once listed, a connector installs in one click across Claude on the web, desktop and mobile, Claude Code and Cowork.

Getting listed in the ChatGPT Plugin Directory

You submit through the plugin portal at platform.openai.com/plugins, and the listing appears in the Plugin Directory that ChatGPT and Codex share. OpenAI asks for more up front:

  • A verified developer or business identity on your OpenAI account.
  • Proof you control the domain: a token served at /.well-known/openai-apps-challenge on the server's host.
  • Three labels on every tool: read-only, destructive, and whether it reaches the open internet.
  • Five test prompts the plugin should handle, and three it should refuse.
  • Public links to your privacy policy, terms, support and website.
  • Reviewer credentials that work without two-factor codes.

OpenAI says review times vary, and nothing is published until it is approved and you choose to publish.

The registry, for everyone else

There is also the official MCP Registry, a public catalog that MCP clients and smaller directories read from. You list a server by publishing one small file describing it. View Printer is there as tech.viewprinter/viewprinter.

Muse, the newest directory

Meta launched Muse, its AI agent app, on September 8, 2026, and opened a connector directory to developers later that month. Applying is one form at muse.ai/platform: you describe the product and how people would use it, Meta reviews it for function, security and legal requirements and tests it end to end, and approved connectors appear in Muse.

Meta's platform page does not say how connectors are built, so whether an MCP server can be listed as it is, or needs work to fit, is still open. We applied with View Printer and will update this section when we know.

What we learned doing it for View Printer

View Printer is in the MCP Registry and in Claude's community plugin marketplace, and its ChatGPT submission went in this month and is waiting on review. A few things we would tell anyone starting:

  • Ship instructions with the tools. Tools say what can be done, not in what order or what to check first. View Printer's plugin carries a skill with nine rules, one for each mistake worth preventing, and each rule is read only when it applies.
  • Words your customers use can mean two things. "Save it as a draft" can mean held back in View Printer or staged on TikTok. The skill makes the assistant ask which, because guessing wrong puts content in public.
  • Some mistakes cannot be undone. A relative time like "tomorrow at 9" passed through without being turned into a date can send a post a day early, and a retry after a lost response posts twice. The rules for those are marked critical.
  • Not every assistant can do everything. Uploading a video takes three steps and a client that can send a file. Where one cannot, the assistant points people back to the website.
  • Sign-in varies between clients. Some assistants support newer ways of registering with your sign-in server than others, so test sign-in in each one you want to be listed in.

Build it yourself, or have it built

Everything above is doable by a team that has built an API before, and the two directories' docs, linked below, are good. If your developers are already committed to the roadmap, we build MCP servers for companies that want their product in Claude and ChatGPT: designed, running in your cloud, and taken through both reviews.

Sources

Checked September 2026.

We help engineering teams get more out of the AI they already pay for. Talk to us about your team.