# What is an MCP connector? A plain-English guide for team leads An MCP connector lets Claude or ChatGPT use another tool on your behalf. What that means, what the sign-in screen grants, and what to ask a vendor first. ## The short answer An MCP connector is a plug-in that lets an AI assistant such as Claude or ChatGPT use another tool on your behalf. You sign in once and approve access; after that, the assistant can call a short list of actions in that tool when you ask. It can only do what those actions allow. MCP stands for Model Context Protocol. It is an open standard, which is why the same connector can work in more than one assistant. You do not need to know how it works inside to make good decisions about it. You need to know what it can do, as whom, and where the data goes. ## An everyday way to think about it Picture a capable assistant who sits next to you all day. Without connectors, they can only work with what you hand them across the desk: text you paste in, a file you upload. With a connector, you have given them a visitor badge for one other office, say the finance system or the project tool, and a short list of errands they may run there. Three things follow from that picture: - **The badge is yours.** The assistant acts as you. It sees what you may see in that tool, and nothing more. - **The errands are fixed.** The connector publishes a list of actions, each with a name, a description and what it needs as input. The assistant picks from that list. It cannot invent new ones. - **You can take the badge back.** Disconnecting the connector ends its access. That is most of it. Everything else is detail about who builds the connector and how it is checked. ## What the sign-in step actually grants When you add a connector, you are usually sent to the vendor’s own sign-in page. You sign in there, with your own account, and approve the connection. This pattern is called OAuth. Its point is that your password goes to the vendor, never to the assistant. The assistant receives a token that lets it act for you within the limits the vendor set. For a team lead, that means: - **Each person should connect with their own account.** A shared admin login would give everyone’s assistant the admin’s view. - **Leaving the company should end access.** When someone’s account in the tool is closed, their assistant should lose it too. Ask the vendor to confirm this. - **The token is not a blank cheque.** It covers the connector’s actions, not everything in the tool. Claude’s directory of connectors makes this a requirement: a listed connector must be a remote HTTPS server that uses OAuth 2.0, and every tool must have a title and be marked as either read-only or destructive. Source: [Claude docs: submitting a connector to the directory](https://claude.com/docs/connectors/building/submission), last verified 2026-10-15. That last mark is worth asking about. A connector that only reads is a small decision. One that sends, changes or deletes something is a bigger one, because a mistake reaches someone else. ## Directory or custom Some connectors are listed in Claude’s Connectors Directory. Directory connectors are available on any paid Claude plan, and Claude can suggest one in a conversation when it fits the task. Source: [Claude docs: Connectors Directory](https://claude.com/docs/connectors/directory), last verified 2026-10-15. Others are added directly by their web address, as a custom connector. The protocol is the same. The difference is how much checking has been done before it reaches you. We cover the organisation settings and a nine-point security list in [Claude connectors for teams](/blog/claude-connectors-for-teams/). ## Six questions to ask a vendor Before a connector goes to the whole team, get written answers to these. Below each one is how Colle answers it, as an example of the kind of answer to expect. ### 1. Where is our data stored? Ask for the country or region, and who hosts it. *Colle:* in the EU (Frankfurt), on encrypted disks with daily backups. ### 2. How is our data kept apart from other customers’? In a shared service, every customer’s data sits in the same system. Ask how one company is kept from reading another’s. *Colle:* row-level security in the database on every query, so one organisation’s request cannot read another’s rows. ### 3. Who inside our company sees what? The connector may store things your colleagues can find. Ask who can see them. *Colle:* work sent to a person is seen only by the sender and the recipients. Work sent to a space notifies its members, and cases in a closed space are visible only to its members. ### 4. Is there an audit log we can export? Ask whether you can see who did what, and when, without asking the vendor. *Colle:* a tamper-evident audit log, linked by a hash chain so a changed or removed entry shows, which administrators can export. ### 5. How does it treat text written by other people? A connector may bring in emails, documents or notes from colleagues. Ask whether that text is handled as data or could be taken as instructions. *Colle:* a colleague’s text is marked as untrusted and scanned, and the assistant is warned. ### 6. How do we sign in, and how do we leave? Ask how people sign in, whether single sign-on is possible, and how you export or delete your data. *Colle:* during the pilot, people sign in through Colle’s own sign-in page with email and password; single sign-on (OIDC) and automatic people sync (SCIM) are set up with you when you need them. An administrator can export or delete company data at any time. If a vendor cannot answer one of these in writing, treat that as the answer. The full list for Colle, including training and the data processing agreement, is on the [security page](/security/). ## How Colle uses a connector Most connectors bring data into the chat. Colle’s does the opposite: it moves work from your chat to a colleague’s. Each person adds it to their own Claude or ChatGPT and signs in. Then: > Send this to James, he runs the launch now The assistant uses Colle’s send action. James is notified at once, accepts, and his own assistant reads the summary, facts and files through the same connector. The inbox and search actions only read; send, return, done, recall and recording a fact change something. Details are on [Colle for Claude](/integrations/claude/) and [Colle for ChatGPT](/integrations/chatgpt/). Because one connector works in both, it is also how mixed teams pass work across; see [teams using Claude and ChatGPT together](/blog/teams-using-claude-and-chatgpt-together/). [Join the Colle pilot](https://www.trycolle.com/#pilot) ## When you do not need a connector at all If your team only asks the assistant questions and drafts text, connectors add little. If the tool you want to reach has no connector, a copied summary still works. Start with one connector that removes a step people repeat every day, and add more only when someone asks for one with a reason. --- Canonical page: https://www.trycolle.com/blog/what-is-an-mcp-connector/ · Last updated: 2026-10-18