Integrations and access

Connect an assistant to approved tools while minimizing access and checking what actions it can take.

Choose the connection that fits

An assistant can work with external systems through built-in connectors, a protocol such as MCP, an API or script, or a browser interface. Which choices are available depends on the product and organization. A connection is not automatically safe or complete: it may expose only some data or actions, and it may behave differently from a person using the interface.

The Model Context Protocol (MCP) is an open standard for connecting AI applications to tools and data sources. A protocol standardizes how clients and servers communicate; it does not certify that a particular server is trustworthy, secure, or authorized to handle your data.

Before connecting a service, ask:

  • Is this integration approved for the data involved?
  • Which accounts, records, and actions will it expose?
  • Can access be read-only or restricted to a narrow workspace?
  • Does the assistant need permission to change data, or only to prepare a draft?
  • How are credentials stored, access logged, and permissions revoked?

Prefer the least access that accomplishes the task. For new or unfamiliar integrations, test with non-sensitive data and inspect the result. Be especially careful with tools that can send messages, edit records, publish content, or trigger other systems.

Handle unavailable integrations

If a prebuilt connector is missing, ask an administrator or developer about an approved MCP server, API integration, or existing automation. Browser-based workarounds can be brittle and should not bypass access controls or organizational review. Avoid pasting secrets into chat or source code.

Exercise: Map one task's systems and data. For each connection, record the minimum required permission and one action that should require human approval.

Sources: Supplied video transcript; MCP introduction; Claude Code MCP guidance