lazysite lets an AI assistant - Claude, ChatGPT, Claude Code, or any MCP client - manage your site through exactly the same rules a person follows. You create a dedicated account for it, grant only the permissions you want, connect the assistant, and revoke or expire access whenever you like. Here is the whole flow.

1. Add a user for the agent

In the manager, open Users and add a new user under your account:

• Choose the AI agent type.
• Give it a distinct name so you can recognise it later - for example chatgpt-web or claude-code.
• Do not assign any groups - the agent gets only the specific permissions you choose in the next step.
• Create it under you, and Add.

2. Choose its permissions

Find the new user under your username and expand its panel. Tick only the permissions you want it to have:

• Manage content - create and edit pages.
• Manage forms - build forms and bind them to delivery.
• Manage themes - change the site's appearance.
• Manage layouts - change the site's structure.
• WebDAV - direct file access, useful for Claude Code and other agentic tools. (WebDAV must be enabled first, in Site settings.)

Grant the least it needs for the job - you can change this at any time (see step 6).

3. Connect the assistant

Open Connect an AI assistant for this user and pick the option that matches your AI. The endpoint is the same for all clients: https://YOUR-SITE/cgi-bin/lazysite-mcp.pl.

Claude.ai (web / mobile)

Keep the Connect panel open. In Claude.ai: Settings → Connectors → Add custom connector; enter your site name and the endpoint URL; Add. Open a chat, enable the connector, and ask Claude to run whoami. When it prompts, paste the connect code from the panel. No token is ever pasted into the chat.

ChatGPT (Plus / Pro / Business / Enterprise)

In ChatGPT: Settings → Apps / Connectors → Advanced → Developer mode → Add custom connector; enter a name and the endpoint URL (OAuth is discovered automatically); create it; then paste the connect code. Plus/Pro can call read-only tools; Business/Enterprise also get the write tools, with a per-call approval.

Claude Code / Claude Desktop / scripts (token)

Use Generate agent brief (a pairing key) or Generate credential (an lzs_ token). Add the endpoint as a remote MCP server with the header Authorization: Bearer <user>:<lzs_ token>. A script can also use the control API and WebDAV directly.

The complete tool list, capability model and error reference is in the AI connector tools reference. If a connection fails, check that WebDAV is enabled and that the connect code has not expired (it lasts 15 minutes - reopen the panel for a fresh one).

4. Put it to work

Once connected, just ask. Depending on the permissions you granted, your AI can:

• Create or edit content - "draft a page about X and publish it".
• Change the appearance - "install the clarity theme" (a theme).
• Change the structure - "switch to a layout with a sidebar" (a layout).

5. Time-box it (optional)

Set the account to auto-expire so access is removed automatically once the work is done - handy for a one-off task.

6. Watch, tighten, and revoke

You stay in control the whole time:

• Did it connect? Check the audit log - every connection and change is recorded, with who did what and when.
• Remove permissions at any time. Once a theme or layout has been built, untick that permission and you know the agent can make no further changes of that kind.
• Disable the account whenever you want to cut access entirely.

Grant narrowly, watch the audit log, and tighten or expire access as soon as the job is done.