How to connect Claude to your documentation (via MCP)

Connecting Claude to your documentation means Claude can search and answer from your actual docs instead of guessing — a big upgrade for any team that keeps real documentation. Here's how to set it up, including the part that matters for client work: doing it safely on private docs.

The mechanism: MCP

Claude clients (Claude Code, Claude Desktop) and other AI clients (Cursor, ChatGPT) connect to external data through the Model Context Protocol (MCP). To connect Claude to your docs, you expose the docs via an MCP server and register that server in your client. Claude then has your documentation as a searchable tool.

Step by step (with Miradorly)

1. Connect your repo to Miradorly (install the GitHub/GitLab app, pick the markdown folders). Your docs are now rendered and indexed.
2. **Add a .docignore** so internal files never reach the index:

CLAUDE.md
**/secrets/**
drafts/

1. Add the MCP server to your Claude client. In Claude Code:

claude mcp add miradorly https://mcp.miradorly.com/<your-workspace>

In Claude Desktop or Cursor, add the same MCP server URL in the client's MCP settings.

1. Ask Claude. It now answers from your documentation — and only from the docs your account is allowed to see.

Why role-aware matters here

If you stop at "Claude can read the docs," you've created a leak risk on shared docs. A role-aware MCP fixes it: the agent authenticates as a specific user (OAuth 2.1 + PKCE) and retrieves only what that user can access. So:

• You connect and Claude sees everything you're allowed to.
• A client connects and their Claude sees only their project's docs.
• Nobody's agent ever sees .docignore-excluded files.

This is what makes connecting Claude to documentation safe to share with clients, not just use internally. More on the concept: what is a role-aware MCP server.

What you can ask once connected

• "Summarize the deployment runbook for project A."
• "What's our policy on X according to the handbook?"
• "Find the API auth section and explain the token flow."

Claude answers grounded in your real docs, with citations to the source pages — within permissions.

Read-only by design

Miradorly's MCP is read-only: Claude can read and answer, never overwrite your docs. Your repo stays the source of truth; the agent is a reader, not an editor.

Bottom line

Connecting Claude to your documentation is two steps — expose the docs over MCP, add the server to your client. The detail that makes it production-safe, especially with clients in the mix, is using a role-aware MCP on private docs so every agent sees exactly what its user can — which is precisely how Miradorly's MCP is built.