Home › Claude Security & Privacy Center for Business Owners › What Claude can and cannot reach once it is connected to your tools

Connectors and data access

What Claude can and cannot reach once it is connected to your tools

Short answer

A Claude connector gives Claude permission to read, and sometimes act in, a specific tool you already run: a mailbox, a shared drive, a calendar, an accounting system. The permissions are set by you, per connector, and can be revoked at any time. The safe pattern for a small firm is read-only by default, one working folder rather than the whole drive, Claude drafts but a person sends, and admin control over which connectors staff may add at all. Nothing Claude reads through a connector is used for training on a business plan, and it is covered by the same Data Processing Addendum as your chats.

How a connector works, and why the plan matters

Claude connects to outside tools through connectors (Anthropic-built, like Microsoft 365 and Google Workspace) and through MCP servers, the open standard Anthropic published for tool connections, which is what Zapier, Composio, n8n and custom-built connections use. In every case the connection authenticates as a specific account with specific permissions, and Claude can only do what that account is allowed to do. On Team and Enterprise, admins decide which remote and local connectors members are allowed to add, which is the single most important security control in a rollout and the one most owners never open.

Claude Organization settings Connectors page with the Let members request connectors toggle and the Add button highlighted above the connector list
Organization settings > Connectors. The admin decides which connectors exist for the organization and whether members can request more.
What each kind of Claude connection can do (summary of Anthropic’s published connector and desktop documentation, checked 2026-11-06)
Connection Reads Writes or acts Scope it
Gmail Read, search, draft Can send, but by default only after you approve each send; owners control this, and my builds keep it draft-only Inbound email can carry instructions Claude might follow (prompt injection)
Microsoft 365 (Outlook, SharePoint, OneDrive, Teams) Read and search what the signed-in user can see Read-only until an admin enables write tools; needs Entra admin consent to connect Grant consent at the tenant level, then decide per tool whether writes are on
Google Drive Read and search files you grant Write back to files only where enabled Scope to a working folder, not the whole drive
Calendar Read, create and change events Invites go to real people Confirm before sending invites outside the company
Accounting, CRM, project tools (via MCP or Zapier/Composio) Whatever the integration exposes Varies; many expose write actions Use a dedicated service account with the minimum role
Desktop app folder access Read and write inside the folders you attach Can create and change files in those folders Attach one project folder; keep client archives out
Browser (Claude in Chrome / built-in browser) Read pages, fill forms, click Can submit forms and complete flows Approve per site; never for banking or payroll

Read-only by default

Most of the value in a small firm comes from Claude reading: the three inboxes, the shared drive, the permit portals, the ticket queue. Writing back is where risk lives. In the builds I run, Claude drafts and a named person sends; Claude updates one workbook the staff member owns and reviews; Claude never submits, signs, pays or sends on its own. One of my engineering clients runs permit tracking this way: Claude reads the agency portals in her browser, read-only, and fills one spreadsheet she checks every morning.

Michael’s two cents

The question I get is “is it safe to connect Claude to our email?” and the honest answer is: it is as safe as the permissions you give it and the inbox you point it at. A read-only connection to a shared inbox is low risk and high value. A send-enabled connection to the owner’s personal inbox is neither. Start narrow, prove the workflow, widen on purpose.

Can an email trick Claude?

“The risk of indirect prompt injection is real and still emerging.”

Max Graupner, CSO Online, Security considerations for adopting Claude Code and Cowork for SMBs, June 19, 2026. Independent source, not Anthropic.

Yes, in principle, and you should plan for it. When Claude reads an email, a web page or a document, that text can contain instructions written to look like they came from you. Anthropic’s products treat tool content as data rather than commands and ask before risky actions, and the desktop app runs file work in an isolated environment, but the owner-level controls are the ones that matter: no send permission on inboxes that receive outside mail, no connector to a system that can move money, and the tool approval mode set so a person confirms anything irreversible. A real exploit against the desktop app was published in early 2026 and patched; it worked because the attacker got a file with hidden instructions onto a machine with broad folder access. Narrow folders would have blunted it.

The desktop app and your files

The Claude desktop app can work directly with files in folders you attach, and only those folders. Attach a project folder, not Documents. Keep client archives, credentials, payroll exports and anything with Social Security numbers out of attached folders entirely. For firms with a file server, Anthropic documents how local file access and the isolated work environment behave; the short version for an owner is that what Claude processes is sent to Anthropic’s systems for the model to read, under the same commercial terms as a chat, so the same data rules apply to a folder as to a paste.

What the admin controls

  • Capabilities: web search, the built-in browser, code execution and its allowed domains, computer access, tool approval modes.
  • Connectors: which remote connectors members can add, whether local (desktop) MCP servers are allowed, and per-product settings for Cowork, Claude in Chrome and Claude Code.
  • Data and privacy: chat and project sharing outside the organization, data export, feedback submission.

Enterprise adds network-level access controls, IP allowlisting, audit logs and the Compliance API, which covers Cowork and Claude Code sessions. On Team, your record of what Claude did through a connector is the connected system’s own log (sent-mail folder, Drive activity, the workbook history), which is usually enough for a firm of 25.

Claude Team admin Capabilities page, Tool approvals section: Automatically approve mode, Skip all approvals mode and Always allow for connector tools
Capabilities > Tool approvals: the switch that decides whether Claude asks a person before using a tool.

Connector questions, answered short

Is it safe to connect Claude to Gmail?

With read and draft permission on a work inbox, yes for most firms. Claude can send mail, but by default it asks for your approval on each send, and owners control that setting. Keep it draft-only unless a workflow needs sending, and never connect an inbox that handles payments without a person approving each send.

Can Claude delete files in Google Drive or SharePoint?

Only where the connection grants write access and the action is permitted. Scope connectors to a working folder and keep archives read-only or unconnected.

Does Anthropic train on data Claude reads through connectors?

No. Connector content on a business plan falls under the same commercial terms and DPA as chats, with training off by default.

Can an admin stop staff from adding connectors?

Yes. Team and Enterprise admins control which remote and local connectors members may add, under Organization settings.

What is MCP and should an owner care?

MCP is the open standard Claude uses to connect to tools. Owners should care about one thing: a custom MCP connection authenticates as an account with permissions, so give it the least role that does the job.

Need a connection that does not exist yet?

Custom Claude connectors (MCP) are built to your permissions: a service account with the minimum role, read-only unless a workflow earns write access, and a log you own.

Custom Claude ConnectorsSee Claude for Teams →

Strategized.ai is an independent Claude consultancy and is not affiliated with, endorsed by, or sponsored by Anthropic. This page explains Anthropic’s published policies for business owners; it is not legal advice. Anthropic’s pages change, so every page in this center shows the date it was last checked.