AI Guides › Claude Mastery
What A Connector Can See (And What It Can't)
By Nigel Guy · 2 min read
The moment people connect Gmail, Calendar or Drive, they tend to swing one of
two ways: either they assume Claude now has full, standing visibility into
everything in their account, checking in on its own — or they assume nothing
has really changed and treat the connection as decorative. Neither is right,
and the gap between what a connector actually does and what it feels like
it does is where most of the confusion (and the occasional genuine surprise)
comes from.
The rule: a connector only sees what you ask it to look up, in that
conversation, at that moment — it is a fetch-on-request tool, not a
background watcher.
The mechanism: the Request/Scope/Duration check
Run this three-part check whenever you're unsure what a connected app
actually exposed:
- Request — did you, or the conversation, actually ask for something
that needed Gmail, Calendar or Drive? If not, nothing was queried. A
connector sits idle until a task calls for it.
- Scope — what was searched or read is bounded by the specific request:
a search for "invoices from March" reads matching messages, not your
entire inbox history, and a Drive lookup reads the files it opened, not
your whole account structure. Broader questions ("summarise my week")
naturally pull a wider scope, and that's worth noticing.
- Duration — access lasts for the conversation's task, not permanently.
Nothing is being monitored between sessions unless you've explicitly set
up something separate, like a scheduled check, and that's a distinct
feature, not a side effect of connecting the app.
Alongside that check, it helps to know what connectors generally cannot do
on their own: they don't send anything without you approving the action,
they don't silently archive, delete or reorganise your files as a matter of
routine, and they don't cross-reference against apps you haven't connected.
What to skip
Skip auditing every connected app "just in case" before each session — if
you're not asking it to touch email, calendar or files, it isn't. And skip
the opposite instinct too: treating a connector as safe to leave switched on
forever without ever checking what it's been asked to do. A quick look at
what a given session actually requested is more useful than either extreme.
Guardrails
- Exact permissions differ by connector and by account type (personal versus
workspace/enterprise), and by what your organisation's admin has allowed —
check your own settings rather than assuming this guide's description
matches your setup exactly.
- What a connector is technically capable of and what it's currently
configured to allow are two different things — read the specific
permission screen when you connect one, don't skim it.
- Anything Claude reads through a connector is still only as accurate as
that source data — a connector can retrieve a wrong or outdated email
just as easily as a right one.
- If sensitive material is involved (financial, medical, legal), treat the
connector the same way you'd treat any other tool with account access:
deliberately, not by default.
All 751 AI guides · JulieMango plans from £17/mo