Glossary
Connector
A connector is a working link between an AI tool and another system, letting it read a shared drive, check a calendar or update a record without anybody copying anything across by hand.
In plain terms
It is a lead between two boxes. One end plugs into the assistant and the other into the system holding whatever you want it to see. What matters about a lead is not that it exists but which way things travel along it and how much it is allowed to touch, and those are the two questions a vendor tends to answer last.
Why it matters
Connector counts are the headline figure in this market and one of the least informative numbers you will be shown. Four hundred connectors is worth nothing if the four systems you actually run are not among them, and one connection to your finance system with permission to write is a far larger decision than four hundred read-only ones. Which systems, in which direction, with whose permissions, and who notices when it stops working: those four questions tell you everything the count does not.
How it works
Underneath it is ordinary software talking to ordinary software, usually through the other system's published interface. What the connector adds is packaging: the sign-in handled for you, the fields matched up, the errors translated into something a person can act on. That packaging is genuinely valuable and it is also why a connector can exist for one product and not for its near-identical competitor.
Every connection has a direction, and products describe both directions with the same phrase. Read-only means the assistant can see things, which is a question about confidentiality. Write means it can change them, which is a question about operations and about what happens on a bad day. A vendor saying a product connects to your systems has told you nothing about which of those you are agreeing to.
Permissions are inherited from the account used to connect, not chosen for the connection. The common mistake is connecting with an administrator's account because it was the one available when somebody was setting it up, after which the assistant can reach everything that administrator can reach. A purpose-made account with only the access the job needs takes ten minutes and is the single most useful precaution here.
Some connectors read live at the moment you ask; others copied everything into their own index and refresh it on a schedule. Those behave very differently when something changed an hour ago, and only one of them will tell you about it. It is worth asking which, because a stale answer that looks live is harder to catch than one that is obviously out of date.
They break, and the breakage is usually silent. The system at the other end changes its interface, a password expires, an approval lapses when somebody leaves. Every connector needs a named owner and a way of noticing failure, because a connection that quietly stopped returning results looks exactly like a system with nothing to report.
The screen everybody clicks past
Seen in the wild
Browse an automation platform's library of pre-built connections and check whether the four systems you actually run are in it, rather than counting the total.
ZapierConnect an assistant to a folder on your own machine and notice what it asks permission to do, which is usually broader than the task you had in mind.
ClaudeUse a workspace search connected to several company systems and confirm that a colleague asking the same question gets fewer results than you do.
Glean
Common misconceptions
People assume
More connectors means a better product.
In fact
It means a longer list. The only ones that matter are the ones you would actually use, and a product with six of those beats one with six hundred of somebody else's. Ask a vendor to demonstrate the specific connection you need rather than accepting the number on the marketing page.
People assume
Connecting something is read-only by default.
In fact
It depends entirely on what was granted, and the grant is frequently broader than the task requires because a broad grant is easier to build against. The screen shown when you authorise it lists the actual verbs. Reading those, rather than the heading above them, is thirty seconds well spent.
People assume
It can see what I can see.
In fact
It can see what the connecting account can see, which is often considerably more. Where somebody connected with an administrator's credentials, an assistant may surface documents to a colleague who could never have opened them directly. This is the readiest way a well-intentioned integration becomes an access problem.
Telling them apart
Connector vs MCP
Connector
An actual working link between two named things. You can switch it on, grant it permissions and watch it fail.
An agreed specification for how such links should be built, so that one of them can serve several assistants.
The lead versus the shape of the plug. These were one entry in the old glossary, which is why they are so easy to run together.
Connector vs API
Connector
Packaged and ready: sign-in handled, fields matched, errors made readable. Built by somebody for one specific pair of products.
The general route underneath, which almost every connector is built on top of, and which a developer can use directly for anything no connector exists for.
If no connector exists for the pair you need, the underlying route usually still does, and the gap between them is a development job rather than an impossibility.
Questions
- How many connectors should a product have?
- Enough to cover the systems you run, which is usually a handful. The total is a marketing figure and comparing two products on it tells you nothing useful. Bring your own list of four or five systems to a demonstration and ask to see each one connected, including the permissions screen.
- Is our data copied, or read at the moment we ask?
- Both arrangements are common and they carry different risks. Reading live means nothing is stored elsewhere and answers are current. Copying into an index means faster searching and a second copy of your material to secure, plus answers that may be as old as the last refresh. Ask which, and if it is copied, ask how often it refreshes.
- Which account should we connect with?
- One created for the purpose, with only the access the task actually needs. Connecting with a personal administrator account is quick, common and the source of most of the awkward discoveries later, because the connection inherits everything that account can reach and keeps it until somebody notices.
- What happens when a connector breaks?
- Usually nothing visible, which is the problem. The assistant carries on answering, simply without whatever that connection provided, and an answer built on missing material looks much like any other. Every connection needs an owner and an alert, and it is worth asking a vendor what their product does when one stops responding.
Key takeaways
- A connector is a working link between two named systems; the count of them tells you very little.
- Direction matters: reading is a confidentiality question and writing is an operational one.
- Permissions come from the connecting account, so a purpose-made account with narrow access is worth the ten minutes.
- Some read live and some search a copy refreshed on a schedule, which behave differently when something just changed.
- They break silently, so every connection needs a named owner and a way of noticing.
Last checked July 2026