How to review AI agent changes to a website before publishing

How an agent reaches your site

Anticode runs an MCP server that assistants connect to with OAuth: Claude (on claude.ai, Claude Desktop and Claude Code), ChatGPT and Cursor have documented setups, and other clients that support remote OAuth servers can connect too. Once connected, an assistant can list and read projects, request changes, read and update project files, start builds, and publish or unpublish sites. Coding agents can also work on the project file directly through the CLI, or by following the open anticode.json format and /AGENTS.md.

That's a lot of reach, so decide what each assistant is for before you connect it. When you approve the connection, Anticode's consent screen lets you grant read-only access instead of read and write; an assistant that only answers questions about your analytics or form submissions doesn't need to edit anything.

Set the publishing rule

Agent drafts, you publish

The safest default for a business site. The agent edits the project; a person reads the preview and presses publish.

Agent publishes small changes

Reasonable for low-risk edits you've described precisely, such as opening hours or a price in a CMS entry, where the previous version is a click away.

Agent publishes freely

Only for sites where a broken hour costs nothing, such as a personal sandbox.

Write the rule into your instructions to the assistant ("make the change and stop; don't publish"). Instructions are a convention, not a lock: an assistant with write access can call publish. If you need a hard limit, connect it read-only, or have it work in the CLI on a copy that a person syncs.

The review, step by step

1. Ask for a summary of what changed. Before you look at anything, have the assistant list the sections and files it touched. A change that wanders beyond what you asked for is the first warning sign.
2. Watch it land in the builder. If you have the builder open, it redraws when the CLI, MCP or an agent edits the project, and offers an undo. Read the changed sections on the canvas.
3. Check on a phone width. Agents often get desktop right and mobile wrong. Read every changed section at phone size.
4. Check the plumbing. In Pro mode, the Map tab draws the site's sections, forms and connectors and checks nothing broke after a change. A rewritten contact section that lost its form shows up here.
5. Run the release checklist. It runs every publish check without publishing, so you see design and security failures before anything goes live.
6. Publish, or send it back. Ask for corrections in plain words, or undo and start again.

What runs no matter who publishes

Publishing through MCP, the CLI or the builder goes through the same server-side gate. The server re-runs the security checks that matter once a page is public: no scripts or inline handlers, no javascript: links, forms that post where they should, HTTPS. A page that fails can't be published. The design gate also blocks pages that miss basics such as a page language, alt text, visible keyboard focus, reduced-motion support, placeholder text or empty links (an empty or "#" link; it doesn't test that outside URLs work).

Those checks catch unsafe or broken pages. They don't catch a wrong price, a deleted paragraph, an off-brand tone or a claim the agent made up. That's what your review is for, and why the Sentinel checks aren't a substitute for reading.

When an edit goes wrong

Open history and preview the version before the agent's change. Each save is a numbered version, and History labels saves made by the CLI or an AI assistant differently from your own, which makes the agent's work easy to spot. Restore the earlier version (the restore is added as a new version, so the later versions stay in the list; only the most recent versions are kept, up to your plan's number), check it, and publish again. More detail is in version history and in making big changes safely.

Limits

• No approval queue. There's no built-in "agent proposes, owner approves" step; the rule is yours to keep, enforced by read-only access if needed.
• You compare by looking. Earlier versions are previewed on the canvas, so reading the changed sections is part of the job.
• Checks are about safety and basics, not about whether the content is true.

Questions people ask

Can an AI assistant publish my site without me?

If you granted it read and write access, yes: publish is one of the MCP tools. Connect with read-only access, or tell it not to publish and review before you do.

Do agent edits skip the security checks?

No. Publishing from MCP or the CLI goes through the server's publish gate, and failing pages can't be published.

How do I undo what an agent changed?

Use the undo the builder offers when it redraws after the edit, or open history, preview an earlier version and restore it, then publish again.