Claude Connectors: What They Are and Which One to Set Up First
A connector is a secure link between Claude and a tool your business already pays for (your CRM, inbox, file store, accounting system), so Claude can read from it, and where you allow it, act in it, without you pasting anything in.
You turn one on in about two minutes: the + button in the chat box, then Connectors, then log in as you would authorise any app. Claude inherits your permissions in that tool, not a master key. There are hundreds available on every plan including Free, and the single most useful thing your team can do this week is connect exactly one.
The copy-paste tax nobody budgets for
Watch how your team actually uses Claude and you will notice most of the effort goes into fetching rather than thinking.
Someone wants a follow-up drafted to a client. They open the CRM, find the account, skim the last three notes, copy them. Then the inbox, find the thread, copy that. Then the proposal in Drive, download, upload. Then, finally, they type the actual request. Four minutes of fetching for thirty seconds of asking.
Multiply that across a team, several times a day, and you have a real and completely invisible cost. But the direct time is the smaller half of it. The fetching quietly decides what people bother to ask. Nobody is going to hand-assemble six months of support tickets into a chat window to test a hunch about churn. So they don't. The question never gets asked and the answer never gets found, and no one ever notices, because an unasked question leaves no trace.
That is what connectors change. Claude is already smart; a connector makes itinformed.
What a connector actually is
A secure, authenticated link between Claude and an outside service. You connect it once by logging into that service through Claude. Two properties matter far more than the rest, and they are the two to say out loud to a nervous team.
Claude inherits your permissions rather than escaping them. If you cannot see a colleague's private folder, neither can Claude working on your behalf. A connector hands Claude your key, not a master key. This one fact defuses most of the anxiety in the room.
Reading and doing are separate permissions. Some connectors only let Claude look. Others let it act, such as sending the email or updating the record. That split is a deliberate control, and on Team and Enterprise plans an administrator can enforce it centrally: connect the tool, restrict Claude to read-only. The jump from reading to doing is the biggest single change in what an AI assistant is, and connectors are where you govern it.
What to connect first
Connect one thing. One connector people actually use beats eight that sit there unopened, and the right first choice is whichever tool your team spends the most time digging around inside.
| Connect | The question it makes possible | Start with |
|---|---|---|
| CRM | “Which deals haven’t moved in three weeks, and what does the last email in each thread say about why?” | Read-only |
| “Read my history with this client, find what we promised in March, and draft an honest status update.” | Read-only, then drafts | |
| File store | “Find every version of this contract and tell me what changed between them.” | Read-only |
| Helpdesk | “What were the five most common complaints last quarter, with a real quote for each?” | Read-only |
| Accounting | “Which costs grew sharply against last month, and is there an obvious reason in the transaction detail?” | Read-only |
| Calendar | “What am I walking into on Thursday, and what do I need to have read first?” | Read-only, then writes |
Look at the helpdesk row. Nobody reads 400 tickets, so that analysis has simply never been done at most companies. It was never hard, just too tedious to be worth anyone's afternoon. Those are the questions that only start existing once a connector does, and they are usually where the first genuinely surprising answer comes from.
MCP, in plain language
You will keep seeing three letters, MCP (the Model Context Protocol), in vendor emails and probably a slide from your IT team. This is all a non-technical person needs to know.
Remember the era before USB. Every device had its own plug; connecting anything to anything was a small research project. Then a common standard arrived and the problem evaporated. AI had exactly that mess: a software company wanting to work with an AI assistant had to build a custom integration, then a different one for the next assistant. Slow, expensive, and the reason your favourite niche tool never had an AI integration.
MCP ended it. Anthropic published the standard openly and it caught on across the industry rather than staying an Anthropic-only thing. A company builds one connector and it works everywhere. That is why the directory went from a handful of integrations to hundreds quickly, and why the specific tool your team actually uses, not just the famous ones, increasingly has one waiting.
The business translation: the plug is standard now, so the remaining constraint is whether anyone at your company has plugged anything in.
To see which tools already have one, and what each lets Claude read and do, browse our directory of MCP connectors.
The two places to be careful
Two, and genuinely no more than two.
Be deliberate about write access. Reading is low-stakes; a wrong summary wastes a minute. Doing is different: a sent email cannot be unsent, and an overwritten record is gone. Start every connector read-only, live with it for a fortnight, then enable writing only where the time saved is obviously worth it. Where you do, prefer Claude drafts, a human sends. That one rule keeps nearly all the speed and almost none of the risk.
Know where your data goes. Data flowing through a connector is handled under that service's terms, possibly in another country. For tools you already use daily this is a non-issue, because you accepted those terms when you bought the tool. But have one short conversation with whoever owns data protectionbefore connecting anything holding personal or regulated data.
What is not worth worrying about, and is where most of the fear actually lands: connectors do not grant Claude some sweeping new access to your company. They grant your access, to one tool you chose, which you can disconnect with a click.
Rolling connectors out to a whole company
Everything above is one person connecting one tool. That is how it starts, and it is also where most companies stall. Six months in, the typical picture: a handful of enthusiasts have connected their own tools, each set up slightly differently, nobody can say which connectors are live, and most of the team has never opened the menu.
Getting from there to a company where connectors are simply how work gets done takes four decisions. None of them is technical.
| Decision | What it means in practice |
|---|---|
| An approved list | Which tools may be connected, decided once rather than negotiated person by person. On Team and Enterprise plans an administrator can switch connectors on for the organisation, so the default is the approved set, not whatever someone happened to find. |
| Read-only by default | Write access granted per tool and per role, only where the time saved is obvious. Sales may draft emails; nobody needs Claude editing the ledger in week one. |
| One owner | A named person who adds connectors when a team asks, removes the ones nobody uses, and answers the question “am I allowed to connect this?” in one message rather than a meeting. |
| A view of what is used | A connector being available says nothing about whether anyone uses it. Look at usage, not licences, or you will not know which connectors earn their place. |
The companies furthest along treat connectors less like apps each person installs and more like shared infrastructure: set up centrally, the same for everyone in a role, with a clear view of what is actually called and by whom. That shift, from personal setup to a managed layer, is where most of the value ends up, because it is what reaches the part of the team that would never have opened the menu on their own. We explain how organisations run that layer in what an MCP gateway is and when you need one, and how to tell whether any of it is working in how to measure AI adoption.
What to do this week
- Connect one tool. The + button in the chat box, then Connectors, then browse. Pick the one your team digs around in most. Log in. It takes about two minutes.
- Ask the question you've been avoiding. Use a real question your team skipped because assembling the data was too tedious, not a demo one. That is the moment this stops being an IT topic.
- Write down the read/write rule before you need it. Agreed calmly now, rather than in a hurry after an AI-drafted email surprises a client.
Copy this into wherever your team keeps its working agreements:
Choosing what to connect, setting the read/write rules, and getting a whole team past copy-paste is most of what a Deployed Kickstart half-day actually covers, mapped to the tools you already run. The Partner programme keeps expanding what is connected, and what your team does with it.
Frequently asked questions
What is a Claude connector?
A secure, authenticated link that lets Claude reach into a business tool you already use (your CRM, inbox, file store or accounting system), so it can read from that tool, and where you allow it take actions in it, without you copying and pasting anything. You set one up by logging into that service through Claude, the same way you would authorise any app.
How do I set up a Claude connector?
Click the + button in the chat box, choose Connectors, browse the directory, pick your tool and log in. It takes about two minutes. Connect exactly one to start with: the tool your team spends the most time digging around inside, usually the CRM, the file store or email. One connector people actually use beats eight that sit there unopened.
Does connecting a tool give Claude access to everything in it?
No. Claude inherits your permissions in the connected service. If you cannot see a colleague’s private folder, neither can Claude when it is working on your behalf. A connector hands Claude your key, not a master key, and you can disconnect it with a click.
Can Claude change things in a connected tool, or only read them?
Reading and acting are separate permissions, and it depends on what you allow. On Team and Enterprise plans an administrator can enable a connector organisation-wide while restricting Claude to read-only. Start every connector read-only, live with it for a fortnight, and where you do enable writing, keep a human between the draft and the send.
What does MCP mean?
MCP, the Model Context Protocol, is the open standard connectors are built on, effectively the USB-C port of AI. Before it, every tool needed a custom integration with every AI assistant. Now a company builds one connector and it works everywhere, which is why there are hundreds rather than a handful, and why niche tools increasingly have one.
Do I need a developer to set up a Claude connector?
Not for anything in the official directory. That is a few clicks and a login, no different from authorising any other app. You would only need technical help to build a custom connector to an internal system that is not listed, and thanks to the shared standard that is now a small job rather than a project.
Are connectors only available on expensive plans?
No. Connectors work across all plans including Free and Pro, and in Claude on the web, Claude Desktop and Claude Code. The administrator controls (connecting tools centrally, enforcing read-only, restricting which services can be connected at all) are the part reserved for Team and Enterprise.
Where does my data go when Claude uses a connector?
To the service you connected, handled under that service’s own terms, which may mean outside your country. For tools your company already uses daily you accepted those terms when you bought the tool. Even so, have one short conversation with whoever owns data protection before connecting anything holding personal or regulated data.
How do we roll out Claude connectors to a whole company?
Make four decisions before anyone connects anything: an approved list of tools that may be connected, read-only access as the default with write access granted per role, one named owner who adds and removes connectors, and a way to see which connectors are actually used. Rolling out without these usually leaves a few enthusiasts with their own setups and most of the team never starting. On Team and Enterprise plans an administrator can switch connectors on for the whole organisation, so the approved set becomes the default.
Can an administrator control which connectors employees use?
Yes, on Claude Team and Enterprise plans. An administrator can switch connectors on for the organisation and restrict whether Claude may only read from a tool or also act in it. That is the practical way to keep an approved list and a read-only default without asking every employee to configure their own setup correctly.
Found this useful? Send it to someone who needs it.