AI Workbench Hub
The hub of an open framework for AI on WordPress: one door for every AI task on a site, and the MCP server AI agents use to work on it safely.
- Part of
- An open framework for AI on WordPress
- Status
- A written design on open standards. Not for sale or download.
- Works with
- WordPress, WP REST API, Abilities API, MCP, OpenAI, Anthropic and Gemini APIs, embeddings
- 01MCPRequestEditor, plugin or agent
- 02AIContextThe site's own content
- 03AIModelOpenAI, Anthropic or Gemini
- 04AIOperationsNo raw markup
- 05ChecksDesign system and role
- 06ReviewA person publishes
How an AI request moves through the hub. A diagram of the design, not a screenshot.
The short version
AI that can't break the site
Changes go through structured operations, checked against the site's design system. AI saves drafts; a person publishes.
One door for every AI task
Editors, plugins and agents all go through the hub, on OpenAI, Anthropic or Gemini, with one set of rules.
An MCP server with limits
AI agents work through an MCP server and get only the abilities their user's role allows.
Grounded in the site
Answers come from the site's own content, so they sound like the site and stay on the facts.
- Open standards: MCP and the WordPress Abilities API
- OpenAI, Anthropic and Gemini behind one door
- Nothing loads for visitors
By the numbers
- 1door for every AI task on the site
- 3model providers, routed per job
- 0lines of markup written by the model
- EveryAI change waits for a person
The problem
A busy WordPress site collects AI fast: one tool for drafts, another for SEO, another for pages, each with its own keys and its own idea of what it may change. Now AI agents want in as well. Without one place that sets the rules, nobody can say what AI is allowed to do on the site.
The pattern
The hub is the site's AI layer. Every AI task goes through it, whether it comes from an editor, a plugin or an agent. It gathers the context the job needs from the site's own content and routes the job to a model from OpenAI, Anthropic or Gemini. Visitors never load any of it.
The point is control. The model never writes markup: it answers in a small set of structured operations, which the hub checks against the site's design system before anything is saved, and it can only save drafts. A person publishes.
AI agents come in through an MCP server on the hub, built on WordPress's Abilities API, and get only the abilities their user's role allows. The site's other plugins plug into the hub too, so none of them carries its own keys or its own rules.
What it replaces
Before
Separate AI tools for drafts, SEO and pages, each with its own keys and its own idea of what it may change, and no safe way in for AI agents.
After
One door and one set of rules: grounded answers, checked operations, drafts for a person, and agents limited to what their role allows.
The gateway
AI Workbench Hub is not one more plugin. It is the door the others go through.
One door
Every AI request on the site, from an editor, a plugin or an agent, enters here and follows the same rules.
The plugins
Form, content, page building, SEO, translation and commerce plugins connect through it, so none of them needs its own keys or its own rules.
AI agents
The MCP server sits on the hub. Each agent gets only the abilities its user's role allows.
AI agents connect through the MCP server and act only within the abilities each user role allows.
AI Workbench Hub vs WordPress's MCP Adapter
WordPress now has a standard way for AI agents to connect: the MCP Adapter exposes a site's registered abilities as tools an agent can call. It is a good bridge. It is not a brain. In the framework, the hub sits behind it and adds the rules.
| Point | WordPress MCP Adapter | AI Workbench Hub |
|---|---|---|
| What it is | A bridge from registered abilities to MCP tools | The site's AI layer, with an MCP server on top |
| Who gets in | Logged-in users, then each ability's own permission check | One door for editors, plugins and agents, with abilities set per role |
| What AI may change | Whatever an ability allows | Structured operations only; the model never writes raw block markup |
| Publishing | Up to each ability | Drafts only; a person always publishes |
| House style | No built-in rules | Every change checked against the site's design system |
| Context | None built in | Grounded in the site's own content |
| Models | None: the agent brings its own | OpenAI, Anthropic and Gemini, routed in one place |
| Other plugins | Not its job | Form, content, page and SEO plugins plug in |
About the MCP Adapter: WordPress MCP Adapter ↗, as of October 2026.
Under the hood
- Request.MCP A request comes in from an editor, from another plugin, or from an AI agent through the site's MCP server.
- Context.AI The hub gathers what the job needs: the page or the brief, and the passages of the site's own content that apply.
- Model.AI The job goes to the model that suits it, from OpenAI, Anthropic or Gemini, with keys and limits kept in one place.
- Operations.AI For page changes the model answers in structured operations, never raw block markup, so what it can do is decided in advance.
- Checks. Each operation is checked against the site's design system and the user's role before it is applied.
- Review. The result lands as a draft. Nothing goes live until a person approves it.
Guardrails
- One door: every AI task on the site goes through the hub and its rules.
- Structured operations only. The model never writes raw block markup.
- Every change is checked against the site's own design system.
- AI saves drafts. A person always publishes.
- Agents get the abilities their user's role allows, and nothing more.
- Nothing loads for logged-out visitors.
Works with
- WordPress
- WP REST API
- Abilities API
- MCP
- OpenAI
- Anthropic and Gemini APIs
- embeddings
What the hub does
Model routing
OpenAI, Anthropic or Gemini per job, with keys, limits and costs in one place.
Grounding
Answers drawn from the site's own pages and documents, so they stay on the facts.
Checked operations
A fixed vocabulary of page changes, validated before anything is saved.
Review queue
Every AI change waits as a draft for a person to approve.
MCP server
AI agents work on the site through the same rules, with abilities set per role.
Why it matters
A site where AI can do real work, for editors, plugins and agents alike, without anyone being able to break it.
What it changes
Editors can use AI on the live site, and nothing goes live without a person.
Every plugin starts with the rules already in place: keys, models, checks and review.
AI agents get a safe way in, limited to what their user's role allows.
6 questions about AI Workbench Hub, answered
What is AI Workbench Hub?
The hub of my open framework for AI on WordPress: one door for every AI task on a site, grounded in the site's own content and routed to OpenAI, Anthropic or Gemini, with an MCP server so AI agents can work on the site safely.
Can AI Workbench Hub break the live site?
No. The model answers in structured operations that are checked against the site's design system, never raw markup, and AI can only save drafts. A person publishes.
How do AI agents connect to it?
Through an MCP server on the hub, built on WordPress's Abilities API. Each agent gets only the abilities its user's role allows, and its changes go through the same checks and review as everyone else's.
Which AI models does it use?
OpenAI, Anthropic and Gemini. The hub routes each job to the model that suits it and keeps keys and limits in one place.
How is it different from WordPress's MCP Adapter?
The MCP Adapter is a bridge: it exposes a site's registered abilities to agents. The hub is what sits behind it: one set of rules, grounding in the site's content, checked operations and a person who publishes.
Can I install AI Workbench Hub?
No. It's a written design on open standards, not a product. The patterns are here for anyone building AI into WordPress.
More plugins
- AI Content PlannerPart of an open framework for AI on WordPress: reads a month of search, analytics and AI visibility data and works with the editor on what to publish next.
- AI ConversionPart of an open framework for AI on WordPress: forms people and AI agents can fill through WebMCP, every submission held until the CRM confirms it, and follow-up drafted for a person to approve.
- AI Visibility TrackerFinds the questions buyers in the company's market actually ask, then measures how the brand shows up when ChatGPT, Gemini, Claude and Perplexity answer them.