Connect an agent

Your agent works on documents through one MCP address of your environment. Every agent takes the same steps:

1. Get a token. Your backend signs it, as Authentication shows. The token names the documents that the agent may reach, and what it may do there.
2. Connect to the MCP address. The client sends the token with each call. Connect your client in the API reference shows how to set this up in Claude Code, Cursor, the Claude Agent SDK, and other clients.
3. Open a session on a document. The agent opens the session itself, or your backend opens it for the agent.
4. Read and edit the document with the tools. The MCP server tells the agent how to use each tool. The tools in the API reference lists every tool.
5. Apply the edits, and close the session. The agent or your backend applies them. Who writes into the document, and when in the API reference explains the options.

How Coauthor works explains what happens in each step.

The type of document changes how the session opens and how the edits are applied. See Document types.

Start with a small job: ask the agent for the headings of a document. The agent opens a session, reads the document, and lists the headings. If it does, the token, the session, and your editor bundle work.

If a call fails, the answer says what went wrong. Errors in the API reference lists every error. If the agent is done but the document has not changed, nobody applied the edits.

• Use cases in the CKEditor 5 documentation show the ways to use Coauthor.
• Security and compliance describes how your documents stay yours, and the guardrails for your agent.
• API reference links to the full reference, and shows how to build an agent that follows changes during Early Access.