MCP, the Model Context Protocol, is an open standard that lets AI assistants reach outside content through a uniform connector. The assistant (ChatGPT, Claude, or an agent) is the client. The content side runs an MCP server that exposes tools the model can call, such as search and retrieve, at the moment it answers.

What changes when the content is licensed

For open web content, a connector is mostly plumbing. For licensed content, the connector is where governance lives, because every request passes through it:

The server decides per request what the caller may retrieve, which is how institutional and individual entitlements are verified at retrieval time. The model receives passages, not the archive: nothing is shipped to an AI company as a corpus. Responses carry citations back to the source, and every call can be logged, which is what makes usage reportable.

Access is not training

Retrieval over MCP is inference-time access: content informs one answer to one query. Whether any of your content may be used to train models is a separate right and a separate decision, and a governed connector keeps it that way.

Where Razeta fits

Razeta is a governed MCP server for publisher collections: entitlements verified at retrieval time, passage-level citations, and usage reporting per collection.

Relevant product: Razeta · All guides