What is an MCP connector?
A connector is how an assistant gets a new ability without anyone rebuilding the assistant. It is also how agents can start sharing context.
Configure · · 4 min read
In short
- An MCP connector is a remote server that an AI assistant connects to in order to use new tools, through the open Model Context Protocol.
- The assistant reads the connector's tool list and decides when to call a tool. The user signs in once and approves access.
- Because the protocol is shared, one connector can work in many assistants without a separate integration for each.
- A connector that reads and writes a profile the user owns lets different assistants share that user's context.
An MCP connector is a remote server that gives an AI assistant new tools through the Model Context Protocol, an open standard for connecting assistants to outside systems. The user adds the connector to their assistant, signs in once, and from then on the assistant can call the connector's tools when a conversation needs them: search a workspace, read a calendar, or save a note.
We build Configure, infrastructure to connect agents, apps and chats, and connectors are one of the ways agents reach it. This is a plain explanation of how they work.
How does an MCP connector work?
- The user adds it. In the assistant's settings, they add a connector by choosing it from a directory or pasting the server's address.
- They sign in. The assistant opens the connector's sign-in page. The user approves, and the assistant receives a token for that user. No key is pasted by hand.
- The assistant reads the tool list. Each tool has a name, a description and the inputs it accepts. The descriptions are written for the model, because the model decides when to use them.
- The model calls a tool when it helps. In the middle of a conversation the assistant sends a request to the connector and gets a result back, which it uses in its reply.
The important design choice is in step 3. Nobody writes code that says "when the user mentions a meeting, call the calendar tool". The model reads what the tool does and works that out. That is why a connector can be added to an assistant its author has never seen.
What is the difference between MCP, a server and a connector?
| Term | What it is |
|---|---|
| MCP | The protocol: the shared rules for how an assistant and a tool provider talk |
| MCP server | A program that exposes tools over that protocol. It can run on your own machine or on the internet |
| MCP client | The assistant or app that connects to servers and calls their tools |
| Connector | What assistants call a remote MCP server once a user has added it to their account |
In everyday use, "connector" means the thing a person turns on in their assistant. "MCP server" is what the developer builds.
Why do connectors matter?
Before a shared protocol, giving an assistant access to a service meant a custom integration for each pair: this assistant with that service. Every new assistant multiplied the work. With one protocol, a service builds a single server and any compatible assistant can use it.
That matters more as the number of agents grows. If we stay on a path where there are more agents than people, nobody can build one-off compatibility with each of them. A common way in is the only approach that keeps up.
What can a connector do, and what should it not do?
A connector's tools can read information, take actions, or both. Good connectors give the user something useful inside the conversation they are having: an answer drawn from their documents, an event added to their calendar, a preference remembered for next time.
What a connector should not be is a way to quietly move a person's data out of an assistant. A tool whose only purpose is to take a user's memory and carry it elsewhere gives them nothing in the session, and it turns the assistant into an exit for personal data. Assistants have good reason to keep that door closed. A connector earns its place by serving the user in context, with their explicit permission. That is the difference between explicit user consent and passive exfiltration.
How do agents share context through a connector?
An assistant's memory normally stops at the edge of that assistant. A connector can change that if its tools read and write a profile that belongs to the user and not to any one assistant.
Here is the pattern. The user adds the same connector to two assistants and signs in to both as themselves. In the first assistant they mention they are vegetarian, and the assistant saves that through the connector. A week later the second assistant, about to suggest a restaurant, searches the same profile and finds it. Neither assistant talked to the other. Both talked to a neutral layer the user controls.
This is how Configure's connector works. It exposes a small set of tools for reading a profile, searching it, saving a fact and forgetting one. Each assistant writes to its own namespace, so the source of every memory is known, and every read is checked against the user's permissions. The user can review, edit or delete anything from their profile. We describe the model in Cross-agent memory: how agents share what they know.
Does every agent need a native connector?
No. A native connector is the smoothest path where an assistant supports one. Beyond that, any agent that can use MCP tools or send an HTTP request can connect, including coding agents, command-line tools and agents with no browser at all. In those cases the agent starts a connection and hands the user a link to approve. The MCP overview covers each route.
How do you add Configure as a connector?
In an MCP-compatible assistant, add a remote server with this address:
https://mcp.configure.dev
The assistant handles sign-in from there. No application code or manual token entry is required. For Claude specifically, see Configure Memory for Claude, which lists the tools and what the connector receives. If you are building your own product and want your agent to use the same tools for each of your users, the quickstart shows how to open a session per user from your server.
Add a connect button to your product
The quickstart takes one page, end to end: a chip in your app, Configure's hosted page, and your agent reading the user's context on the first turn.
Read the quickstartNext: AI onboarding without the questionnaire