Claude connectors and MCP, explained

On its own, Claude can only talk. It can explain how to deploy a website, but it cannot deploy one. It can draft a reply to a support ticket, but it cannot see the ticket. Connectors close that gap. Once one is added, Claude can reach into that service while you chat: read what is there, and act on it when you say so.

This guide is for people who use Claude rather than people who write software. It covers what connectors and MCP servers are, how to add one, what they are good for, and how to stay in control of what Claude does with them. We make a connector ourselves, so there is a worked example near the end. Everything before it applies to any connector.

What a Claude connector is

A connector is a link between Claude and another service: your notes, your calendar, your code, the place your apps run. It gives Claude a set of actions in that service, such as "search the documents", "create an issue" or "deploy this app", and Claude decides which to use based on what you ask.

You talk to Claude the way you already do. Ask "what's on my calendar tomorrow?" with a calendar connector added, and Claude looks it up rather than telling you it cannot. Ask it to "put the meeting notes in our team wiki" and, with a wiki connector, it does that too.

Anthropic has also used the name integrations, and the services themselves often call them MCP servers. All three describe the same thing, seen from different sides.

What an MCP server is

MCP, the Model Context Protocol, is an open standard Anthropic published in November 2024 for connecting AI assistants to other software. An MCP server is the piece a service runs so that assistants can use it. It publishes a list of what it can do, and for each action, what information it needs and what it returns.

Because it is a standard, one MCP server works with any assistant that speaks MCP. The same server you add to Claude can usually be added to Cursor, VS Code, ChatGPT or your own agent, and the service builds it once rather than once per assistant.

Most connectors you add to Claude are remote MCP servers: the service runs them, and you add them by their web address. Claude's desktop app can also run local MCP servers on your own computer, which is mostly useful to developers. This guide is about the remote kind.

How to add a connector to Claude

There are two ways, and both take about a minute.

From the directory. Claude has a directory of connectors from companies that have published one. Open it, find the service, and add it. This is the easiest route, and listings there are reviewed by Anthropic.

By its address. If a service is not in the directory, or your company's workspace limits the directory, go to Settings › Connectors › Add custom connector and paste the address the service gives you. It looks like any web address, often ending in /mcp.

Either way, the first time you use it Claude sends you to the service to sign in. You approve what Claude may do there, and you are sent back. You do not copy or paste any keys, and you can remove the connector whenever you like from the same settings page.

On a company plan, an administrator may need to allow a connector before people in the workspace can add it. Which connectors your plan allows is in Anthropic's help pages.

What people use them for

The useful ones are the services you already switch between all day. A few of the common patterns:

• Find and summarise. "What did we decide about the launch date? Check the meeting notes." Claude searches the notes and answers with where it found it.
• Hand over the busywork. "Turn this list of bugs into issues and assign them to me." What was ten minutes of copying becomes one request.
• Run a workflow across tools. "Check this week's sign-ups, and post the three biggest accounts in the sales channel." Two connectors, one request, and Claude does the steps in order.
• Ship what you built. If you build apps with Claude, a hosting connector lets Claude put the app online, give it a database and tell you when it is live, without you leaving the chat.

The last one is where connectors change the most for people who are not developers, so the example below uses it.

A worked example: from a chat to a live app

We host apps and run an AI gateway, and our connector lets Claude do both from a conversation. This is what using it looks like.

1. Add it. Open Roar AI in Claude's connector directory and add it, or paste https://cloud.roar-ai.com/mcp as a custom connector.
2. Approve it. Claude sends you to Roar AI to sign in, or to make an account. A page names the app asking for access, the account it will act as and what it will be allowed to do, and you choose Allow or Cancel.
3. Ask for what you want, in plain language.

Some of the requests people send:

• "Deploy the intranet repository from GitHub with a database, and tell me when it is live."
• "Why did the last deploy of the support bot fail? Show me the build log."
• "What did we spend on AI this month, and which app spent the most?"

Claude picks the right actions, runs them, and reports back with a live address or the exact error. Each action tells Claude whether it only reads or changes something, and which changes cannot be undone, and Claude uses that to check with you before it acts.

Some things are deliberately not offered to an assistant at all, because a person should decide them: inviting a colleague, adding credit, editing the rows inside an app's database and restoring a backup are done in the console.

Which connectors to add first

Start with the two or three services you open most, not with everything in the directory. A good first set for most teams:

• Where your documents live: notes, wiki or shared drive, so Claude can find what you already wrote.
• Where your work is tracked: issues, tasks or a CRM, so requests turn into real items.
• Where your apps or data live, if you build things: hosting, a database, error tracking.

Prefer a service's own connector to one made by someone else, and check what it asks permission for before you approve it. A connector that only needs to read should not be asking to delete.

Using the same connectors in Claude Code

Claude Code, the version of Claude that works in a terminal, uses the same MCP servers. You add one with a single command and the same address:

claude mcp add --transport http roar https://cloud.roar-ai.com/mcp

Replace the name and the address with those of the service you are adding. The first time Claude Code uses it, you sign in through your browser, as you do in the Claude app.

Staying in control

A connector acts as you, with the access you approved, so the same care applies as with any app you sign in to:

• Give the least access that does the job. If a service lets you choose what Claude may do, choose the narrowest option.
• Read before you approve. Claude asks before a change that matters. Look at what it is about to do, especially anything that deletes.
• Remove what you no longer use. In Claude, Settings › Connectors, then remove it, and Claude stops using it straight away.

Connectors make Claude far more useful than a chat window on its own, and most of that comes from a small number of well-chosen ones. If building and shipping apps is part of your work, our connector is one to try: add it, ask Claude to deploy something, and see the address it hands back.