WordPress AI agent
A WordPress AI agent that works inside rules your site enforces.
A WordPress AI agent built on GoBuildWP connects ChatGPT, Claude, Cursor and other MCP clients to typed WordPress abilities, while your site decides the authority, holds what needs a person, and records what changed.
Free plugin, no account, 26 client guides, Pricing
Representative session. Ability names and behaviour match GoBuildWP Free 1.18.6; the site is an example.
Definition
What a WordPress AI agent is, and is not
A WordPress AI agent is an AI client that operates a WordPress site through a defined set of operations, instead of writing text for a person to paste. With GoBuildWP those operations are abilities registered on the site through the WordPress Abilities API and served over MCP.
The model still reasons and proposes. What changes is where authority lives: on the site, enforced per call, whatever the prompt said.
| Question | Typical AI writing plugin | Agent through GoBuildWP |
|---|---|---|
| What it produces | Text, usually inserted into an editor | Typed operations on the site: pages, products, settings |
| Who decides authority | The person who accepts the text | The WordPress user, the profile and per-ability switches |
| Before a change | Nothing to review but the text | For supported post, media and settings projectors, a diff you can apply or discard in wp-admin |
| After a change | A post revision, if any | A redacted ledger entry and rollback where supported |
Three kinds of agent
Which WordPress AI agent do you mean?
Hosts, chatbot plugins and MCP servers all sell a WordPress AI agent. They differ on where the model runs, what it may touch, who says yes to a risky change and how you take one back.
| Agent | Runs where | What it can change | Who approves | Undo and record |
|---|---|---|---|---|
| Host agent: AI Agent by SiteGround | A chat in wp-admin that calls SiteGround's AI Studio API; installable on any host | Posts, pages, WooCommerce products and orders, plugins, themes, users and settings | Power Mode, off by default, must be switched on before deletions, bulk and settings changes | WordPress post and page revisions; an activity log records page and post changes |
| Host agent: Hostinger Agent (Kodee) | A chat in wp-admin on Hostinger WordPress hosting plans | Settings, users, posts, pages, media metadata, WooCommerce products, stock and orders | The agent asks permission before it modifies the site | Not documented on Hostinger's page |
| Chatbot plugin: AI Engine | Workspace chat in wp-admin, plus iOS and Android apps. | Read and write tools are separated; admins only unless editors are allowed. | A person approves inside AI Engine's Workspace; external clients are left to confirm. | Not documented by AI Engine. MCP logs with per-call metrics and CSV or JSON export. |
| MCP agent from your own client, through GoBuildWP | Claude, ChatGPT, Cursor or another MCP client you already use, calling your site's own endpoint | Only the abilities the connected user, the safety profile and your per-ability switches allow | Destructive calls need confirmation, which the agent supplies itself by default; Free's A person approves setting requires a person, and Pro's opt-in policy can hold ordinary writes as well | A redacted change ledger per agent credential, with rollback checked against the before-image where a restore is supported; irreversible changes are marked as such |
Read from each vendor's own pages on 29 September 2026. A host agent is the quickest start on that host; an MCP agent keeps the client you already use and puts the rules on the site. AI Engine compared in detail
One controlled loop
Discover, inspect, review, execute, verify
The agent should use the live site's schemas and return evidence. It should never invent availability, grant itself access, or treat a confident answer as proof of a WordPress or provider outcome.
A site can require preview before every write that supports it. The setting is off by default because it changes the contract for connected agents. Unsupported projectors are unaffected, and built-in Chat has its own approval path. How preview works
Discover
List the abilities the current site and identity actually expose. Anything the profile removed is simply not there.
Inspect
Load the exact schema and the current object before planning a change: a page snapshot, an element tree, a product.
Review
For a supported projector and readable target, preview-ability returns a field-by-field diff and a wp-admin link. Review unsupported fields, side effects and warnings too.
Execute
Submit the typed operation. The profile, capabilities, confirmation rules and any enabled Pro approval policy decide whether it runs.
Verify
Read state back and check what the site served, plus any browser or external provider evidence the change depends on.
The boundary
What the agent is refused, and by whom
Every call is decided by the WordPress user behind the credential, the active safety profile, the abilities you enabled and the confirmation gates, on the server.
- Read Only
- Discovery and reads. The right profile for the first connection of any client and for any agent you have not yet watched work.
- Production Safe
- The working profile. Refuses raw PHP, WP-CLI, filesystem and database access, plugin and theme installation and temporary administrator access.
- Developer Full Access
- For deliberate staging work. The usual mistake is switching to it once and leaving it there.
- On every profile
- Destructive and critical calls need an explicit confirmation, writes are rate limited per credential, and single abilities can be switched off.
- #1520Update post, Contact (applied from preview)Rollback available
- #1519Delete the old Offers draftConfirmed first
- #1516Update post, Offers (draft)Rolled back
Pro When the approval policy holds a write for this agent, the call returns a pending response, so a well-behaved agent reports the change as not yet made.
The ledger records which credential acted, which ability ran and when. Values under password, token, API key and cookie fields are redacted; other inputs are stored truncated, so the ledger can still hold site content. Ledger and rollback
Free foundation, Pro depth
The control layer first, then the plugins you run
An agent is only as useful as the operations it can call, and only as safe as the policy around them.
The MCP server, client setup, WordPress identity, discovery, safety profiles, per-ability controls, confirmation, rate limiting, preview before write, the change ledger and rollback. Posts, pages, media, menus, comments and native Elementor content work.
89 plugin-aware integrations across builders, forms, SEO, custom fields, WooCommerce, multilingual, events, memberships and learning, each loading only when its plugin is active. Plus approvals, memory and integration health.
FAQ
What a controlled agent really means
What can a WordPress AI agent do?
It can only call abilities registered and allowed on the live site. GoBuildWP Free provides the MCP foundation and controls; Pro adds plugin-aware modules. Availability also depends on the connected user, the profile, installed plugins and the licence.
Is a prompt enough to keep an agent safe?
No. Prompts are context, not authorization. WordPress capabilities, server-side safety profiles, per-ability controls, confirmations, rate limits and approval gates have to enforce authority outside the model.
How is this different from an AI writing plugin?
A typical writing plugin calls a model and puts text into an editor for a person to review and save. GoBuildWP registers typed abilities on your site and lets an external client call them under WordPress capability checks, so the agent operates the site rather than drafting text for it. That is why authority has to be enforced server-side.
Which AI clients can connect?
GoBuildWP's Connect screen generates setup for 27 clients, and 26 of them have a setup guide on this site. The authentication route varies by client; use the live guide and generate credentials on the site rather than copying a generic secret.
Does it need a paid AI subscription?
GoBuildWP does not include a model. Your connected client supplies model access, and the optional in-dashboard Chat uses whichever provider the WordPress AI Client is configured with.
What happens when the agent gets something wrong?
Supported operations are written to a redacted change ledger on your own site, recording which credential acted, which ability ran and when. Retained recovery data and a supported restore are required. Inspect list-sessions with the session_id for conflicts, unchecked targets and not_reversible or not_redoable items; unsupported entries refuse a session run unless allow_partial is true. A failed restore stops the run, and a successful rollback updates the original record. It is not a backup: a delivered email, a fired webhook or a CDN copy cannot be recalled.
What are the requirements?
WordPress 6.9 or newer and PHP 8.0 or newer. Authentication is OAuth 2.1 with PKCE or a WordPress Application Password. A hosted client such as ChatGPT also needs the site on public HTTPS; a local client can use the mcp-remote bridge, or, on the machine that runs the site, WP-CLI over stdio.
Start with a smaller, observable agent.
Connect one client on staging, begin in Read Only, and prove a reversible task before granting broader authority.