How to connect a knowledge base to Claude via MCP

To connect a knowledge base to Claude with MCP, you add the knowledge source as an MCP connector in Claude's settings, using the endpoint your provider gives you. Once connected, Claude can query that source directly during a conversation, with nothing to paste or upload.

What MCP actually does here

The Model Context Protocol is an open standard that lets AI assistants call external tools and data sources during a conversation. For a knowledge base, that means Claude can send a query to your connected source, a document set, a database, or a curated brain, and get back the relevant passage in real time as part of answering your question.

Without MCP, the alternatives are manual: paste the content, upload a file, or maintain a Claude Project. MCP replaces that manual step with a live connection.

The general setup

The exact screens change as Claude's interface evolves, but the shape of the process stays the same.

• Get the MCP endpoint from your knowledge provider. With KBrain, the endpoint is generated automatically when you create or subscribe to a brain.
• Add it as a connector in Claude's settings, under connected MCP servers or connectors.
• Authorize the connection if the source requires it. Private brains usually do, marketplace brains may not.
• Query normally. Ask something the connected knowledge covers, and Claude retrieves from the source while it answers instead of leaning on training data alone.

What changes once it is connected

• No more re-pasting. The knowledge persists across every conversation, not just the one where you uploaded a file.
• Retrieval, not context-stuffing. Claude pulls the specific relevant passage instead of you loading a whole document into the chat.
• One connection, current knowledge. When the underlying source updates, a synced folder or a maintained brain, Claude sees it on the next query with nothing to re-upload.

Common issues

• Claude does not seem to use the connected knowledge. Check that the connector is enabled for the conversation. Some Claude interfaces require you to toggle connectors on per chat, not just connect them globally.
• The connection needs authorization you do not have. Private brains are scoped to whoever created them. To query someone else's private brain you have to be granted access first. That is by design.
• Answers still read generic. Confirm the question falls within what the connected brain covers. MCP retrieval only helps where the source has an answer. Outside that scope, Claude falls back to its own reasoning.