Security for agent integrations

What the server enforcesCopy Anchor Link

• No privileged path. A token is one person in one organization, with exactly that person's access. An agent cannot do anything its person could not.
• Admins and members only. Learners and requesters (people whose request to join is pending) cannot authorize, and the role is checked again on every call. See Authentication.
• Organization-bound tokens. The organization is fixed at authorization and survives workspace switches in the app. Crossing organizations needs a new authorization.
• Short-lived, audience-bound tokens. Access tokens last one hour and are accepted only by the MCP endpoint they were issued for.
• No writes over MCP. Every tool is a read. Building, editing, and publishing happen in the Vidext app, where publishing a course is always a person's action.
• Bounded reads. Knowledge is read through opaque handles that search returns; arbitrary paths and altered handles are refused.

Practices for your sideCopy Anchor Link

• Verify the endpoint. The official MCP URL is https://app.vidext.com/api/mcp/operator. Be suspicious of anything else claiming to be Vidext.
• Use clients you trust. The client sees every tool result, including excerpts from your organization's sources.
• Mind prompt injection. Agents that read content can be steered by it. Treat search results and source text as untrusted input when deciding what your agent does next.
• Least privilege. A member's token reads everything an admin's does over MCP, so there is no reason to connect with an admin account.
• Offboarding. Removing someone from the organization stops their calls immediately. Also remove the connection from their AI clients; Vidext has no in-app list of connections to revoke.