Skip to content
Call wiki

How does Claude actually see your Drive, Slack, and email?

Last updated 2026-08-17 · Overview — no new claims; each statement links to the entry that carries the primary sources.

Connectors are how Claude reaches your Drive, Slack, email, and calendar. Four things about them surprise almost everyone.

A connector is a live search, not a copy

Section titled “A connector is a live search, not a copy”

Nothing is uploaded or indexed in advance. Each time Claude needs something, it fires a query at the service right then, built from whatever keywords it picked (Doc freshness). So connector reads are always current — and only as good as one keyword search. Files you upload by hand are the opposite: snapshots that never refresh.

It sees the whole account

Section titled “It sees the whole account”

A connector sees everything the connected account sees — the entire Drive, Slack DMs included — and telling it to stay out of a folder is a request, not a boundary (Connector scope). What actually scopes access: a second, limited account; per-tool read/write permissions; org-wide settings on Team plans. Each Google connector holds one sign-in, and a second replaces it (Two Google accounts). What I say on calls: decide the boundary yourself — connect work sources freely, and keep the truly sensitive source (patient data, your personal notes) in an account Claude was never given.

Don’t trust it — test it

Section titled “Don’t trust it — test it”

Agents report success whether or not the connector worked; a failed web fetch can silently turn into an answer built from search snippets (Web fetch limits). Before building on any connection, I make it prove itself: fetch a known document, pull the last four Slack messages, quote a page verbatim. Then keep it honest by asking for tweet-length verbatim excerpts with links, not summaries.

Search is shallow — give it a map, or the files

Section titled “Search is shallow — give it a map, or the files”

Drive search through the connector is a literal keyword query returning about ten results, blind to most non-Google file types (Drive search). Two fixes, both of which I set up on most calls. First, a navigation map: have the agent crawl the Drive once and write down which folder holds what — the Drive can stay perfectly disorganized. Second, localize: download a project’s documents into a local folder, where access is certain and search is exhaustive; the same move works for whole websites (Site corpus). Prefer text formats — CSV over .xlsx (Feeding spreadsheets), markdown over Docs (Markdown vs Office).

Know who owns what it reads

Section titled “Know who owns what it reads”

On a Team plan the organization owns the data: the Primary Owner can export your conversations, though teammates see nothing and Anthropic doesn’t train on any of it (Team account data). A personal account is the mirror-image problem — no processor contract covers other people’s data you put there (GDPR personal accounts). The rule I give: work data on the work account, personal life on the personal one, and change your behavior rather than connect a source that mixes them.

No connector? There’s still the browser

Section titled “No connector? There’s still the browser”

A browser agent can do anything clickable through your signed-in Chrome, contained by site permissions (Browser agents). It’s the slow, last-resort route — use it for sites nothing else reaches, not as a substitute for a connector that exists.

Go deeper

Section titled “Go deeper”