Connect Letter.md to AI with MCP

Letter.md provides a remote MCP server for bringing your documents into AI workflows. Connect once, then let supported clients read the document you choose, understand your current selection, and create changes you can review.

Your document does not change when an AI proposes an edit. Suggestions and comments remain reviewable until you accept them.

1. Add the remote MCP server

In ChatGPT, Codex, Cursor, or another client that supports remote HTTP MCP and OAuth, add this server:

https://mcp.letter.md/mcp

The client should discover Letter.md’s OAuth configuration automatically.

Cursor

Add the following to your mcp.json:

{
  "mcpServers": {
    "letter.md": {
      "url": "https://mcp.letter.md/mcp"
    }
  }
}

2. Sign in and authorize

The first connection opens Letter.md sign-in. Sign in and approve access. The authorization belongs only to that MCP client, and you can remove the Letter.md connection from the client at any time.

3. Open a document and collaborate

Open a Letter.md document in the MCP App editor, then ask the host to review it, comment on it, or suggest edits. Review suggestions one by one and accept only the changes you want.

Permission scopes

• documents:read — find and read documents you can access
• suggestions:write — create reviewable suggestions and AI comments
• documents:write — create or save unpublished documents in writable Workspaces when you explicitly request it; existing documents are not overwritten
• assets:write — upload generated images to editable documents

Letter.md requests the smallest scopes required by the workflow. Connecting one client does not authorize another client.

Troubleshooting

• Confirm that your client supports remote HTTP MCP and OAuth.
• Use the MCP endpoint ending in /mcp; the editor URL is not the server address.
• If authorization is stuck, remove the old Letter.md connection from the client and add it again.
• If a document is unavailable, confirm that the signed-in account has access to its Workspace.

WebMCP in supported Chrome

On supported Chrome builds, Letter.md also exposes page-scoped tools through document.modelContext. Open Letter.md normally; no remote MCP address or separate OAuth connection is required.

With the document open, an agent can read the complete Markdown, outline, or current selection; find and open documents; inspect document assets; search Assets and collections; and create reviewable suggestions. Creating a document or suggestion always asks for confirmation.

Available tools follow the current page, Workspace membership, and document role. Direct edits, asset deletion, publishing, and sharing are not exposed as WebMCP tools.

To inspect the tools registered for the current page, open Chrome DevTools and choose Application → WebMCP. WebMCP is still an experimental browser capability; unsupported browsers continue to work normally without these tools.