What are agent connectors
Last updated
An agent connector is how an AI agent reaches a system outside the conversation. The agent needs a defined way to search, read, or change that system. The connector supplies it. It can be a Model Context Protocol (MCP) server, an HTTP API the agent calls, or a plugin that exists inside one product. The job is the same: turn an outside capability into a tool the agent can invoke, and return a result the agent can use.
This directory lists MCP servers. Each listing is one server, with a public source, a short description taken from that source, and JSON and Markdown copies. A page here does not run the server and does not score it.
What an MCP server is
MCP is an open protocol. A host, such as an editor or a chat application, speaks MCP to a server. The server names the tools it offers. The host calls a tool with arguments, and the server does the work: query a database, call an API, read a document, or drive a browser. The protocol is the message format between host and server. It does not host those servers.
An MCP server is one kind of agent connector. Connector and plugin are everyday names for the same idea. The listings here are MCP servers from public sources, including the official MCP Registry and the Model Context Protocol servers repository.
APIs and plugins
Not every connector speaks MCP. A product may only offer an HTTP API, reached through a client or a small wrapper. A plugin may exist only inside one application. Those are still connectors. The practical questions are which host you use, whether it can talk to the connector, and what the connector is allowed to do once connected.
How to choose
Start with the job. Search, databases, browser automation, payments, and documents are different jobs, which is why the listings are grouped into categories. Open a category, then open a listing and read the source it cites.
Publisher. A vendor listing is tied to the organization that owns the product. A reference listing comes from the modelcontextprotocol organization. A community listing is published by a community, contrib, or ecosystem project. Unknown means those checks did not confirm another class. The class is a label, not a ranking, and not a statement that the server was tested.
Remote or local. A remote server has a network address. The host connects to it, and that server sees the requests you send. A local server runs on your machine: the host starts a command, and files or credentials stay there unless the server sends them on. The listing records the form the source documents. If the source does not say, the page does not guess.
Auth. Some servers need an API key or OAuth. Some need nothing. Connect notes on a listing repeat variable names the source already publishes. They are not a new requirement. Keep a key out of chat logs, and give a remote server only the access the task needs.
Using this directory
Browse by job from the category list. Search and retrieval, data and databases, and browser and automation are three of the groups. Every listing page names its source URL, and has JSON and Markdown copies.
To search the catalogue from an agent, the connect page documents a free, read-only MCP server on this site. It returns listing text. It does not install, host, or run the servers in the results. How a listing is added or left out is explained on the about page.