Connector

An integration that exposes selected data or operations from another service to an AI product, usually through authenticated tools and provider-defined permission scopes.

A connector links an AI product with another service such as email, a calendar, a document store, or a meeting recorder. Authorization commonly uses the provider's sign-in flow and returns a token with named scopes. The connector then exposes operations the runtime can call, such as listing events or creating a document. Its actual reach is the intersection of the provider scopes, connector implementation, account policy, and tools the agent receives; the label alone does not establish those limits.

Builder example

A connector changes both available context and possible side effects. An assistant can summarize calendar events with read access, while writing the result needs a separate destination capability. Choose the smallest scopes and operations that complete the job, separate read and write paths when possible, and test the provider's real enforcement. Product interfaces and scope choices change, so record what was granted and provide a way to revoke it.

Common confusion: A read-only provider scope should prevent writes through that credential, even when the model requests one. It does not prevent the workflow from using a different write-capable connector, cached credential, or application endpoint. Evaluate the complete capability graph rather than assuming one consent screen governs the whole agent.