I wanted people to be able to let an agent make changes in a real CMS without giving it unrestricted control - and without first installing WordPress, creating a staging site, or configuring a local environment.
So I built a live playground that runs WordPress inside the browser and lets an external agent work through a controlled MCP interface.
The result is here:
- Playground overview: https://wpsuite.io/agent-playground/
- Live sandbox: https://preview.wpsuite.io/
- Agent Composer: https://wordpress.org/plugins/smartcloud-agent-composer/
The interesting part is the editing boundary.
A real WordPress sandbox in the browser
The playground runs a real WordPress instance in WebAssembly.
You choose one of four prepared site presets and start a disposable WordPress session. There is no VPS to provision and no local WordPress installation to prepare, the browser becomes the temporary WordPress host.
That makes the demo easy to start, but it creates another problem: an external MCP client such as ChatGPT or Codex cannot simply connect to a server listening inside a browser tab, so the browser needs a bridge.
The MCP relay
The playground uses a session-scoped relay between the external MCP client and the WordPress sandbox.
At a high level, the path looks like this:
ChatGPT, Codex, or another MCP client
|
| MCP
v
session-scoped relay
|
| browser connection
v
WordPress in WebAssembly
|
v
WordPress MCP Adapter
|
v
Agent Composer
|
v
validated Gutenberg draft
The relay solves transport.
It does not decide what the agent is allowed to do.
That boundary stays inside WordPress.
Composer is the control layer
Agent Composer is not an AI model. It is a governed execution layer for agent-assisted WordPress editing.
A site administrator can define versioned Config Sets with Site Contracts and page-type Blueprints. These can describe approved structures, patterns, fields, relations, media rules, and other constraints.
Composer then exposes only its governed operations through the WordPress MCP Adapter.
This is very different from giving an agent broad WordPress administration access and hoping the prompt is strict enough.
The agent can work only through the operations Composer exposes.
Draft-first by design
The most important rule is simple:
The agent drafts. Humans publish.
Agent-facing Composer operations create or update Composer-owned drafts, they do not publish normal site content. They also do not provide a general content deletion path.
That makes the workflow useful for real editorial work because the final approval stays in the normal WordPress review process.
Validation matters more than generation
The playground demo uses ChatGPT, but the model is replaceable.
In the demo video (see the cover), I attach a short product brief for Composer Pro and ask ChatGPT to create a new feature draft.
The brief is written before the agent starts. The agent does not invent the product from a one-line prompt.
Composer provides the site context and controlled actions. It then validates the draft against the active rules.
The output is still native WordPress content.
It is not an exported screenshot, a proprietary page format, or an opaque AI-generated site.
Why Gutenberg is useful here
Gutenberg is not everyone's favorite editing experience, that is actually one reason this workflow is interesting.
An agent can do the first structured pass while the result remains a normal Gutenberg draft.
A human can then open the editor, change blocks, rewrite sections, move content, replace media, or reject the draft completely.
The agent reduces the blank-page work without replacing WordPress as the editing surface.
Why build the playground at all?
Agent workflows are difficult to explain with architecture diagrams alone.
A live disposable sandbox makes the behavior visible:
- You can see the agent inspect the site context.
- You can see which Composer actions it calls.
- You can see the draft appear in WordPress.
- You can then open the result as normal WordPress content.
That is a much better test than another AI-generated landing page screenshot.
Try it
The playground is live here:
The overview and instructions are here:
https://wpsuite.io/agent-playground/
Agent Composer itself is free and available from WordPress.org:
https://wordpress.org/plugins/smartcloud-agent-composer/
The project is still evolving, and I am especially interested in feedback on the execution model.
