Umbraco has spent the last couple of releases baking AI directly into the backoffice — suggestions that appear right where you're already working, whether that's tightening a meta description, drafting alt text for an image, or translating a field into another language. Tools like this Editor MCP server take a different approach entirely: rather than living inside a single field, they hand an AI assistant the same access a developer would have to the whole site — every page, every language, every media item — and let you ask for what you want in plain English. Having spent the last couple of days testing the MCP route hands-on against a live Umbraco site, the honest answer to which is "better" is: it depends what you're trying to do.
Where backoffice AI wins
Umbraco AI's advantage is that it never leaves the page you're editing. There's no assistant to explain your site structure to, no picking the right page out of a tree, no risk of the tool touching something you didn't mean it to — the suggestion appears against the field you're already looking at, and you accept it, edit it, or ignore it. For genuinely single-field tasks — rewrite this paragraph, suggest alt text for this one image, translate this label — that's a shorter, safer round trip than typing a request to a chat assistant and waiting for it to find the right page first.
Where an MCP-connected assistant wins
An MCP-connected assistant earns its keep the moment the job stops being about one field. Over the course of testing, a single conversation added German and Greek translations to all twenty dictionary items, audited and fixed SEO metadata across eight blog pages, and pulled a full publish-date report across the whole site — the kind of work that, at the field level, means opening the same screen twenty separate times. Because the assistant is calling the same underlying API a developer would use, it can also chain steps together: create a page, set its author and image, then publish it, all from one instruction, rather than requiring you to know which screen each step lives on.
The tradeoff: wider reach, wider risk
That wider reach comes with a wider blast radius, and it's worth being honest about it. During testing this week, one tool call reported success — a valid confirmation, a generated ID, all the signs of a completed action — while quietly writing nothing at all, because the block type it targeted didn't match what the page actually used. A separate call failed outright with a raw server error rather than a helpful message, because a picker field expected a value shape the tool's own documentation had wrong. An in-context suggestion in the backoffice can't fail that way: it's validated by the same form you'd use to enter the value yourself, so a bad result is visible on the spot rather than buried inside a batch of twenty.
Access itself is a different kind of dependency, too. Umbraco AI in the backoffice is authenticated for exactly as long as you're logged into the CMS — nothing else to think about. An MCP server sits between you and the site as a separate piece of infrastructure, holding its own token to the backoffice on your behalf, and that token can expire or need refreshing independently of your own login — something this testing ran into directly, with every tool suddenly returning an unauthorised error mid-session for no visible reason. None of that is a reason to avoid the approach, but it is a reason to treat it as another system to keep an eye on, not a feature that simply works because Umbraco does.
So, which is better?
Put together, this doesn't look like a contest with one winner so much as two tools built for different scopes of problem. Quick, contained edits are better served by AI that stays inside the field you're already in — there's less that can go wrong, and less to explain. Work that spans many pages, many languages, or several structural changes at once is where an MCP-connected assistant earns its place, provided the server itself gets more disciplined about surfacing failures honestly rather than reporting success when nothing was actually written. For a team weighing this up, the more useful question probably isn't "which AI is better" but "how much of my site does this task touch" — and choosing the tool that matches.