Claude Connectors: What They Are and What Claude Can Actually See

Most people use Claude by copy-paste: paste text, get an answer, copy it back. Connectors remove that step - they link Claude to services you already work with, such as Google Drive or Slack, so it can search, read and act directly against your real files.

What they actually are

A connector is an authorised link between Claude and an external service. Once connected, you can ask for things involving your real content - find a document in Drive and summarise it, pull data from a spreadsheet, send a Slack message - without pasting anything.

They work in several places: Claude in the browser, Claude Desktop, Claude Code and through the API. And there is a Connectors Directory showing what exists - each with a page detailing what it can read and what it can change.

Under the bonnet they are built on MCP, an open protocol for connecting AI assistants to tools. You do not need to understand it to use them, but two consequences are worth knowing: the list keeps growing, and any service supporting MCP can connect - including an internal system of your own. See also what MCP is.

The permission model - and this is the part that matters

This is the first question people ask, rightly. Three facts:

1. Claude inherits your permissions rather than gaining new ones

It cannot reach what you cannot reach. If you have no access to a particular Drive folder, you will not get to it through Claude either. The connector mirrors what you already have - it does not open a new door.

The practical implication matters: if you connect an account with broad permissions, that is what becomes available. In an organisation it is worth considering connecting with an account whose permissions match what is genuinely needed.

2. It accesses when asked, not continuously

Claude reaches your data when your question or request requires it - it does not scan in the background. Ask a general question unrelated to your files and there is no reason for it to touch them.

3. Actions that change things ask for approval

And this is the most important distinction:

Before connecting, open the connector's page in the directory and see exactly what it can do. "Connected to Drive" can mean read-only or full file management, and that is a difference worth knowing in advance rather than discovering.

What they are good for

• Finding and working on a document without hunting for it and downloading it.
• Summarising scattered material - several documents on one subject.
• Pulling data from a spreadsheet and analysing it.
• Cross-checking - what a document says against what a spreadsheet holds.

What they share: you are present, asking for a specific thing, and getting an answer. That is the work they help with.

Where this is not the answer - the distinction that saves a project

This is the point most worth understanding before getting excited.

A connector is assistance, not automation. It helps you do a task while you are there. It does not:

• Run at 3am with nobody asking
• Respond automatically to every incoming lead
• Sync two systems in the background every five minutes
• Run a process that must happen identically every time

Why that matters: those are precisely what people picture when they hear "Claude is connected to my systems". And those are automation - needing a trigger, fixed logic, error handling, and a process that runs without you.

The practical distinction:

And the two are complementary: plenty of businesses use a connector for daily work and build automation for whatever must run unattended.

Before connecting - four checks

1. Which account are you connecting? Its permissions are what becomes available.
2. Read only, or write too? Check the connector's page rather than assuming.
3. Who else has access to this conversation? On a team account that is a separate question.
4. Is the information sensitive? Customer data, medical or financial information - that is an organisational policy question rather than a technical decision, and in a regulated business it is worth settling before rather than after.

That last point matters particularly in Israel: if the information includes customers' personal data, deciding what may be passed to an external service is a question for a lawyer, not for whoever sets up the connection.

If there is no connector for what you need

Three options, in order:

1. Check the directory again. It grows, and common tools are usually already there.
2. Check whether the service has its own MCP server. Increasingly they do - see what an MCP server is.
3. Build a connection. That is development work, which pays off when the need recurs rather than when it is one-off.

But before the third option, worth asking: is the real need genuinely "Claude should see the system", or is it automation someone described in Claude's vocabulary? If the latter, a connector will not solve it.