How do you connect Claude to your company's data? SharePoint, Google Drive, Slack, databases and ERPs

You connect Claude to company data through connectors. Use the ready-made ones in the Connectors Directory for Microsoft 365, Google Drive, Gmail, Slack and many SaaS tools. For databases, ERPs and internal systems, point a custom connector at a remote MCP server you build or buy. An Owner adds each connector for the organization, then each person signs in with their own account, so Claude only sees what that person can already see.

This guide covers the decision for each kind of system, the exact admin steps in Claude, powered by Anthropic, and the governance points that matter most. Product details are as of October 2026 and come from Anthropic's help center and connector documentation, listed at the end. For the protocol itself in plain terms, see our explainer on MCP and Salesforce.

What is a Claude connector?

A connector is a link between Claude and an MCP (Model Context Protocol) server, which a service runs so Claude can search it, read from it and, if allowed, act in it. Connectors show up in Customize > Connectors in claude.ai and the desktop app, and the same catalog serves claude.ai, Claude Desktop, mobile, Claude Code and Cowork. That is different from uploading a file: a connector reads current data when someone asks a question that needs it, under that person's permissions, instead of a static copy someone dragged into a chat.

Connectors come in three kinds. Remote connectors, the large majority, connect from Anthropic's cloud to a service on the internet and work on web, desktop and mobile. Local connectors are desktop extensions packaged as MCP Bundles that run on one computer and work only in the desktop app. Custom connectors are any remote MCP server you add by URL. In the directory, each listing is labeled Verified (Anthropic tested its tools for quality and compatibility) or Community (screened, not reviewed in depth). Anthropic is explicit that verification is not a security audit, and that the developer controls the tools and can change them after review.

Which connection method fits each system?

How do you add a connector for the whole organization?

On Team plans, Owners and Primary Owners add connectors. On Enterprise, so can anyone with a custom role that includes managing Libraries. For a custom connector, the help center gives this sequence:

• Go to Organization settings > Connectors.
• Click Add, hover over Custom, and select Web.
• Enter a name and the remote MCP server URL, then click Continue.
• Review the detected authentication settings and click Continue.
• Set Authentication to Sign in now, Sign in when needed, or No sign in.
• Choose the OAuth client. The recommended option uses the published Claude identity; the alternatives are Register automatically and Use your own OAuth client.
• Add any fixed credentials, such as an API key or bearer token, under Request headers. They are stored encrypted and sent with every request.
• Click Add. Members then open Customize > Connectors and click Connect.

Directory connectors follow the same pattern without the URL step. On Team plans, members without permission see a Request button, and requests land in Organization settings > Connectors under Requested by your team and in Organization settings > Notifications. To change a custom connector's details later you remove it and add it again. Remote connectors call your server from Anthropic's cloud, so a private or firewalled server needs Anthropic's IP addresses allowlisted. The Microsoft 365 article gives the range as 160.79.104.0/21, which also matters if you use Entra Conditional Access.

Per-user sign-in or a shared credential?

This is the most important governance decision in the whole setup, and it is easy to get wrong because the shared option is quicker to build.

When do you need a custom MCP server, and what must it support?

Build or buy a custom MCP server when no directory connector exists for the system, when the existing one exposes more than you want, or when you need business logic between Claude and the data, such as returning a summarized ledger rather than raw journal lines. Anthropic's builder documentation sets these requirements and limits:

• Transport: Streamable HTTP. The older HTTP+SSE transport still works but is being deprecated.
• Authentication: OAuth 2.0 per user, a static credential an Owner enters once, or none. Claude supports the 2025-03-26, 2025-06-18 and 2025-11-25 MCP authorization specifications and can register itself through Dynamic Client Registration.
• Redirect URI: allow https://claude.ai/api/mcp/auth_callback for the hosted apps, plus a loopback redirect for Claude Code.
• Size and time limits: about 150,000 characters per tool result and 240 seconds per tool call in claude.ai and Desktop; 25,000 tokens per result in Claude Code by default, configurable with MAX_MCP_OUTPUT_TOKENS.
• Not yet supported by the Claude client: resource subscriptions, sampling, and draft capabilities. Do not design around them.
• Testing: you test against the real Claude client by adding the server as a custom connector, and can validate auth with the MCP Inspector.

Design tools around questions people actually ask, not around tables. A tool called get_open_invoices_by_customer with a 200-row cap is safer and more useful than a run_sql tool. Return the fields people need, enforce row limits in the server, and log every call with the user identity.

How do you reach systems inside your private network?

You have three options. First, give the MCP server a public endpoint and allowlist Anthropic's IP range at your edge. Second, on the Enterprise plan, request access to MCP tunnels, which are in research preview. You run a small stack of two containers (Cloudflare's cloudflared and Anthropic's mcp-proxy) inside your network; it opens an outbound-only connection on port 7844, so no inbound firewall rules are needed. Limits: up to 10 active tunnels per organization, a 90-day server certificate, IPv4 only, and by default only RFC 1918 private addresses. The tunnel carries traffic but does not authenticate to your servers, so each MCP server still needs OAuth. Third, a local desktop extension runs on a user's machine and uses the local network, but it does not work in claude.ai or Cowork.

A tunnel belongs to the organization that created it. A tunnel created in claude.ai cannot serve the Claude API, and one created in the Claude Console cannot serve claude.ai. If you plan custom agents later, see our guide to the Claude API versus Managed Agents. Note that the API's MCP connector is not eligible for zero data retention.

What security controls should be in place before you connect sensitive data?

• Clean up source permissions first. Claude mirrors each user's existing access; it does not fix oversharing, it makes it easier to find.
• Keep an approved-connector list and route new requests through IT. Prefer Verified listings and connectors run by the system's own vendor.
• Start read-only. Turn on write tools for a small group first; on Enterprise, Microsoft 365 write tools can be limited to some users through custom roles.
• Disable tools you do not need from the Search and tools menu, and reserve Allow always for trusted servers and tools.
• Disable write-capable tools before Research, which can call connector tools without asking each time.
• Review requested OAuth scopes during sign-in, and watch for tool behavior that changes after a server update.
• Train people on prompt injection: text inside a document or email can try to instruct Claude. Our guide on whether MCP is safe covers the controls in more depth.

What is a sensible order for connecting systems?

• Step 1, IT and security: one document source, Microsoft 365 or Google Drive, read-only, after a permissions review.
• Step 2, IT: email and calendar for the pilot group, still read-only.
• Step 3, department owners: two or three directory connectors tied to named workflows, such as the ticketing or project tool a pilot team uses daily.
• Step 4, data team: one custom read-only MCP server over the warehouse, with curated tools and row limits.
• Step 5, security: write tools for a small, trained group, with audit log review.
• Step 6, infrastructure: tunnels or ERP connectors once the earlier steps have run cleanly.

Connectors give Claude access; skills teach it how your team uses that access. Pairing each connector with a short skill, for example how to read your chart of accounts or which SharePoint library is authoritative, is often the difference between a demo and daily use. Plugins can bundle the two so people install both at once. Abstrakt Solutions builds connectors, custom MCP servers and the governance around them; see our Claude connectors and MCP services.