Connect an AI assistant to Docly with MCP

What you'll see

You want Claude, ChatGPT or another MCP client to read and edit a Docly workspace, and wonder where to point it, how it signs in, and why it says it needs approval before it can write.

What's actually happening

One address. Add https://cloud.docly.net/mcp as an MCP server (a remote, or “custom”, connector). The client sends you to Docly's sign-in, which requires two-factor authentication, and then to a consent page. The tools reach the same functions as the docly CLI, so the account needs developer access in Docly.

The consent page decides the binding:

• What the connection reaches: one workspace, or all workspaces you have access to. Everything outside the binding is refused, reading included.
• Whether it may write. For one workspace, this is decided here.
• How long the connection lasts: 12 hours, 30 days, 90 days or one year. The client renews its access in the background for as long as the connection lasts.

Writing with “all workspaces”. The connection can read everything you can read, but it writes only where you have approved it. When the assistant tries to write in a workspace without approval, it gets a link of the form https://cloud.docly.net/mcpwrite?r=… and is told to wait. You open the link, signed in, and approve writing in that workspace for 1 or 8 hours. The approval belongs to the connection, so it holds across chats that use it.

What MCP does not do. It cannot publish or unpublish, and it cannot change shares; it can only show them. docly unlock opens folders for the CLI only — it has no effect on an MCP connection, and an MCP approval has none on the CLI.

What to do

• Bind a connection to the one workspace the work is about, and give it write access only when it needs it. Prefer a short lifetime for a connection that can write.
• See and remove your connections, approvals and pending requests under Connected apps in your Docly account (https://cloud.docly.net/connectedapps). Disconnecting takes effect on the next call.
• To change the workspace or the write access of a connection, connect again.
• Treat a write-approval link as a request from the assistant: approve it only when you expected it, and check that the workspace it names is the one you meant.
• Give the assistant the knowledge base, as for the CLI: How an AI agent should use the docly CLI covers the same rules.