Skip to main content

Connectors (MCP) in the Apps

Intermediate
What you'll learn
  • Understand what a connector adds that the chat alone cannot do
  • See why connectors, Claude Code MCP servers and API MCP are the same standard wearing three hats
  • Set one up with least privilege, in three steps
  • Know the two risks you are accepting: access, and untrusted content arriving through it

Connectors let the Claude apps reach outside the chat — into your tools and data (drives, issue trackers, calendars, databases, and more) — so Claude can answer from, and act on, real systems. Under the hood they're powered by the open Model Context Protocol (MCP).

What they do​

Without connectors, Claude only knows what's in the conversation. With a connector, it can (with your permission) pull relevant info from a connected service — e.g. find a document, read recent issues, check a calendar — and use it in its answer.

Same standard, everywhere​

Connectors are the app-facing form of MCP. The very same protocol powers MCP in Claude Code and on the API. Learn the concept once; it applies across surfaces.

Set up & use​

Guided walkthrough1 of 3
  1. Authorize via OAuth, where supported. This is the step that creates the trust relationship — everything after it inherits whatever you grant here.

Ask across a connected service

Find my Q3 planning doc and summarize the risks.

Safety​

A connector is access + (sometimes) actions
  • Authorize only services and scopes you trust.
  • Content pulled from external sources can carry prompt injection — be cautious when a connector reads untrusted material.
  • Review what a third-party connector can do before enabling it (Reviewing Third-Party Code).

Check yourself

0/3
  1. A connector pulls in a shared document that someone outside your team wrote. What is the specific risk that creates?
  2. You only need Claude to find and read documents. What scope should you grant?
  3. What is the relationship between app connectors, MCP servers in Claude Code, and MCP on the API?
Key takeaways
  • Without a connector Claude knows only what's in the conversation; with one it can read from and act on real systems, with your permission.
  • Connectors are MCP wearing its app-facing hat — the same protocol as Claude Code MCP servers and API MCP.
  • Setup is three steps: connect, grant least privilege, then just describe the goal.
  • You are accepting two risks: the access you granted, and untrusted content arriving through it. Scope narrowly and treat external material as untrusted.

Next​