Breakdance 3.0: MCP-Ready AI Agents Inside Your Page Builder

Breakdance 3.0 does something no other mainstream WordPress page builder shipped first: it hands the builder to an AI agent over the Model Context Protocol. You point Claude Code or Codex at the site, describe what you want, and the agent builds it — not as an exported HTML blob, but as real, editable Breakdance elements.

That is a genuinely new capability. It is also in beta, and the way it authenticates deserves more scrutiny than the launch coverage has given it.

TL;DR

• What shipped: Breakdance 3.0, adding MCP support so AI agents can build directly in the builder.
• Version status: Public beta as of July 28, 2026 — Beta 1 on July 20, Beta 2 on July 26. Stable line is still 2.8.1.
• Agents supported: Claude, Codex, Cursor, OpenCode, and in principle any MCP client. Connection runs through the Agent Connector plugin.
• What an agent can build: Pages, templates, headers, footers, selectors, CSS variables, responsive styles across Breakdance breakpoints, dynamic data via ACF and Meta Box, Posts Loops with conditional display, and multi-field forms with actions and redirects.
• Also in 3.0, and not AI: Redesigned Selectors system with no automatic .breakdance wrapper, native CSS variables, a new Interactions system, and builder performance work.
• The catch: Connecting an agent means issuing a WordPress application password with administrator-level access.
• Do today: Install it on staging, not production. The selector specificity change alone justifies a test pass.

How the MCP integration actually works

The setup path is Breakdance → Settings → Agents & MCP. From there Breakdance offers a one-click install of Agent Connector, a companion plugin that exposes the site to MCP clients. You pick your agent from a list, generate an application password, copy a command into your terminal, and confirm the MCP server registered.

The interesting design decision is what happens after that. The agent does not manipulate Breakdance's internal data structures directly, and it does not drop a block of static markup into a page. It writes HTML and CSS, and Breakdance converts that output into native elements and selectors.

That conversion step is the whole ballgame. It means:

• Agent-built pages open in the visual builder as normal, editable elements.
• A client or junior team member can adjust an agent-built page without knowing it was agent-built.
• The agent reuses your existing selectors and CSS variables instead of inventing a parallel set of one-off styles.

The last point is what separates this from the "AI website builder" features bolted onto other builders over the last two years. Those typically generate a self-contained chunk of markup with inline styles that fights your design system. Breakdance's approach maps agent output onto the design system you already have.

An agent can reach further than page layout, too. It can build Posts Loops with conditional display logic, wire dynamic data through ACF or Meta Box, and construct multi-field forms with configured actions and redirects — the parts of a build that normally involve the most repetitive clicking.

The part nobody is highlighting: it's an admin credential

Here is the sentence from the connection flow that matters most to anyone managing client sites: you generate an application password, and that password grants administrator-level access.

Not builder-scoped access. Not a read-only inspection token. Administrator.

An MCP client holding that credential can do anything an admin can do through the WordPress REST API — read user data, install plugins, change settings — because application passwords inherit the full capabilities of the user that issued them. WordPress has no mechanism for scoping an application password down to "Breakdance layouts only." That is a WordPress limitation rather than a Breakdance bug, but the practical exposure lands on you either way.

Agent Connector gives you three controls worth using from the start:

• Disable the connector when you are not actively building. There is no reason to leave the door open between sessions.
• Lock it to a specific domain, so a leaked credential is less portable.
• Log MCP events, which gives you an audit trail of what the agent actually did.

The risk profile here is not "the AI will go rogue." It is the ordinary one: a long-lived admin credential now lives in an MCP client config file on a laptop, and possibly in a dotfile that gets synced or committed. Treat it like any other production admin secret, because that is what it is.

The non-AI changes are the ones that might break things

The MCP feature is opt-in — it does nothing until you install Agent Connector. The rest of 3.0 is not opt-in, and one change carries real regression risk.

Selectors no longer get an automatic .breakdance wrapper

Breakdance 3.0 ships a completely redesigned Selectors panel, and selectors it generates no longer sit inside an automatic .breakdance wrapper class.

This is a CSS specificity and cascade change, not a cosmetic one. The old wrapper added a class's worth of specificity to every builder-generated rule. Custom CSS you wrote to override builder styles was calibrated against that specificity. Remove the wrapper, and the balance of power shifts — overrides that used to win may now lose, and rules you wrote defensively at high specificity may now override things you did not intend to touch.

Sites most exposed are the ones doing the most hand-written CSS: heavily customised agency builds, anything with a large stylesheet of overrides, anything where a developer fought the builder to get a layout right. Sites built with builder controls alone will mostly be fine.

Native CSS variables

3.0 adds first-class CSS variable support for colours, fonts, and spacing, usable across elements and selectors. This is straightforwardly good and overdue — it is how you build a maintainable design system, and it is what lets an agent reuse your tokens rather than hardcoding hex values.

Interactions

A new Interactions system handles click, hover, keyboard, and scroll-based behaviour, with the ability to chain multiple actions and have Breakdance run them in sequence. Keyboard triggers being included is a nice accessibility signal — most builder interaction systems are mouse-only and force you into custom JS for keyboard parity.

Performance

Broad builder-speed improvements: page loading, element rendering, and dynamic data processing. Worth noting these are builder performance gains — editor responsiveness, not front-end Core Web Vitals.

Should you install it?

Not on production. Not yet. 3.0 is two weeks into public beta, Beta 2 is two days old at the time of writing, and it changes CSS generation. Those three facts together make a straightforward case.

Yes on staging, this week, if you run Breakdance seriously. Two reasons. First, the selector specificity change is something you want to discover on a clone rather than on a client's live homepage, and you need lead time to plan those fixes. Second, the MCP integration is worth forming a real opinion about early — agent-driven building is going to be a competitive factor in how fast agencies ship, and the gap between "read about it" and "know what it's actually good at" is wide.

A sensible test sequence:

1. Clone a representative site to staging — pick your most CSS-heavy build, not your simplest.
2. Update to 3.0 and diff the rendered pages against production. Look specifically for spacing, typography, and anything you remember fighting.
3. Only then install Agent Connector, using a dedicated throwaway admin user.
4. Give the agent a real task you can verify — rebuild an existing page you know well, so you can judge output quality against a known-good reference.
5. Check what it produced in the builder, not just in the browser. Editable native elements are the entire promise; confirm it holds.

How this compares to the rest of the field

Breakdance is not the only builder in the Soflyy stable to get MCP this month — sibling product Oxygen shipped its own MCP support in 6.2, with an overlapping feature set and the same Agent Connector plumbing. Given the shared codebase lineage, that is unsurprising, but the licensing models diverge sharply: Breakdance is $199.99/year, Oxygen is a one-time lifetime licence.

Against the wider market, Bricks Builder has no comparable MCP integration as of late July 2026, and Elementor's AI features remain generation-oriented rather than agent-oriented — you prompt for a section, you do not hand an agent the keys to the site. Breakdance's bet is that the useful unit of AI assistance is a connected agent with real context about your design system, not a text box that emits a hero section. On the evidence of 3.0, that bet looks correct, even if the implementation needs another few releases to earn production trust.

Related reading

• Breakdance vs Bricks Builder 2026
• Best WordPress Page Builders 2026
• WordPress 7.1: AI Client, Streaming and Semantic Search
• How to Build a WordPress Site With AI in 2026
• WordPress Staging Site Guide
• WordPress Security Hardening Checklist