NK · Nov 10, 2026 · 19 min read

You have probably used an AI chatbot to draft a paragraph and then pasted it into WordPress by hand. That works, but it keeps the assistant outside your site. It cannot see your posts, check your plugins or fix the typo on page 14 unless you copy everything back and forth. An AI agent on WordPress changes that. It connects to your site, reads what is there and, within limits you set, makes changes for you.

This guide walks you through your first AI agent on WordPress from start to finish. You will understand what the connection is made of, create a safe user and an application password, connect an assistant, run a first harmless task and then decide what to allow next. Set aside about 20 minutes if you already have a site and an AI assistant account. The whole way through, the advice is conservative: staging first, least permissions, review before publishing.

You do not need to write code. You do need admin access to a WordPress site, and you should be ready to use a copy of your site rather than the live one for the first run.

Key takeaways
  • An AI agent on WordPress is an assistant that talks to your site through a standard connection (MCP), using abilities that WordPress or plugins expose.
  • Use a dedicated WordPress user with the lowest role that does the job, and an application password you can revoke, never your main admin login.
  • Start on a staging copy. Treat the live site as the second step, not the first.
  • Begin with read-only tasks such as audits and lists, then move to drafts, and only later to publishing or deleting.
  • You stay responsible for the result. Keep backups, read the changes and revoke access when you stop using the agent.

What an AI agent on WordPress actually is

A chatbot answers questions. An agent takes actions. The difference is tools. When you give an assistant a set of tools, it can decide to call one, read the result and carry on. On WordPress, a tool might be “list the latest posts”, “update this post”, “check which plugins have updates” or “create a draft page”.

Three pieces make this possible:

  • The assistant. This is the AI app you talk to, such as Claude, ChatGPT or a coding assistant. It decides which tool to use.
  • The connection. This is the Model Context Protocol, usually shortened to MCP. It is a shared standard that lets an assistant discover and call tools on another system. If you want the plain-language version, read what MCP is and why WordPress cares.
  • The tools on your site. In WordPress these are called abilities. A plugin or WordPress itself registers an ability, such as “create post”, with a description of its inputs and a permission check.

So the chain is: you ask the assistant, the assistant calls an ability over MCP, WordPress checks that the user behind the connection is allowed to do it, and then the action runs. The permission check in the middle is the part that keeps this safe, and it is why the user you connect with matters so much.

How the pieces fit: Abilities API, MCP Adapter and application passwords

If you only want the steps, skip ahead. If you want to know what you are switching on, this section is worth two minutes.

The Abilities API

WordPress added the Abilities API as a central registry where core, plugins and themes declare what they can do in a form that both people and programs can read. Each ability has an ID, a label, a description, a category, an input and output description, a function that does the work and a permission callback that decides who may run it. We explain this in more depth in the Abilities API explained. The key point for a beginner is that abilities are described and permission-checked, which is much safer than handing an AI a raw database login.

The MCP Adapter

The MCP Adapter is an official project that bridges the Abilities API to MCP. According to its repository, WordPress abilities are private by default, so you have to opt an ability in before it can be used as an MCP tool. Once exposed, abilities show up as MCP tools, resources or prompts. The default server provides three generic tools: one to discover abilities, one to get information about an ability and one to execute an ability. The adapter supports an HTTP transport on a REST route and a STDIO transport that runs through WP-CLI.

You do not need to install the adapter by hand in every setup. Some plugins and hosts ship their own MCP server that works on the same ideas. For example, AI Engine’s page says its MCP server lets agents such as Claude and ChatGPT read and edit a site through many tools, with undoable changes and permission-aware access. NibWP, which WPMusk uses, is another example; see our NibWP review for how it is positioned. Which one you choose matters less than understanding the safety model described below.

Application passwords

Application passwords are a WordPress feature that lets a program log in as a user without knowing that user’s real password. You create one per application, from the user’s profile screen, give it a name and copy the generated string once. You can revoke it at any time without changing the user’s login. It carries the permissions of the user it belongs to, so a password created for an administrator can do anything an administrator can do.

WordPress expects application passwords to be sent over HTTPS. If your site is not on HTTPS, fix that before you connect anything. Some setups also use OAuth instead, which several guides describe as the preferred path where a tool supports it. If your tool offers OAuth, it is worth reading its documentation. This article uses application passwords because they are the most widely supported starting point.

For a wider view of how this fits into the rest of your stack, the complete guide to AI and MCP on WordPress covers the whole picture.

Before you start: the 20-minute checklist

Gather these before you touch any settings. Most failed first attempts come from missing one of them.

  1. A WordPress site you can use for practice. A staging copy is ideal. If you do not have one yet, the staging site guide shows you how to make one. A throwaway install on a local tool also works.
  2. A recent backup. If you do not already run backups, see the backup plugin comparison. Make one before you start, even on staging.
  3. HTTPS on the site. Application passwords and tokens should travel over an encrypted connection.
  4. A current WordPress version. The Abilities API arrived in WordPress 6.9, and the MCP Adapter targets that version or later. Check the current requirements in the project repository, because they change.
  5. An AI assistant that supports MCP. Claude Desktop, Claude Code and ChatGPT with connectors are common choices. Your plan or workspace may limit custom connectors, so check that first.
  6. Node.js, if your setup uses a local proxy. Many guides use a small local package that turns local MCP calls into WordPress REST calls. Others connect directly over HTTP and need nothing installed.

Time check. Backup and staging are the slow parts. If those already exist, the rest takes under 20 minutes.

Step 1: Create a dedicated user with the lowest role that works

Do not connect the assistant as your own admin account. Create a separate user for it. This gives you three benefits. You can tell in the activity log what the agent did. You can limit what it is allowed to do. And you can delete the user to cut access instantly.

  1. In wp-admin, go to Users, then Add New.
  2. Choose a clear username, for example ai-agent-staging, and use a real inbox you control for the email.
  3. Pick the role. The right role depends on the first task:
    • Editor is enough for reading and editing posts and pages.
    • Author or Contributor is enough if you only want drafts that someone else reviews.
    • Administrator is needed for plugins, themes and settings, which you can add later and only on staging at first.
  4. Set a long random password and store it in your password manager. You will not type it again, but the account still needs one.

Here is the principle to hold on to: the agent can do what its user can do, and nothing more. If you start with a Contributor and the agent tries to publish, WordPress itself will refuse. That is the safety net doing its job. Raise the role only when you have a reason.

Step 2: Generate an application password

  1. While logged in as an administrator, open the new user’s profile (Users, then Edit).
  2. Scroll to the Application Passwords section.
  3. Enter a name that tells you what it is for, such as Claude desktop on laptop.
  4. Click Add New Application Password.
  5. Copy the password immediately. WordPress shows it once. Store it in your password manager, not in a chat window or a note on your desktop.

If you do not see the Application Passwords section, one of three things is usually the cause. The site may not be on HTTPS, a security plugin may have disabled the feature, or your host may have turned it off. Fix the cause rather than working around it, because the feature being absent is often a deliberate setting.

One password per device or tool is a good habit. If a laptop is lost, you revoke that one password and everything else keeps working.

Step 3: Install or enable the MCP side on your site

Your site needs something that speaks MCP. You have three common routes. Pick one.

Route A: The official MCP Adapter

This suits developers and anyone comfortable with Composer, plugins or WP-CLI. You install the adapter, make sure the abilities you want are registered and mark them as public for MCP. Because abilities are private by default, nothing is exposed until you say so. That is the correct default, and you should keep it that way for anything that changes data.

Route B: A plugin that ships an MCP server

Some plugins add their own MCP server and tools through a settings screen. AI Engine, for instance, lists a built-in MCP server in its free version according to its plugin page, and Elementor’s plugin page describes an Elementor MCP aimed at building pages. This is the quickest route for beginners, but read what the plugin exposes. A plugin that offers a hundred tools is powerful and has a larger surface to review. Our AI plugin roundup and plugin vetting checklist help you judge it.

Route C: A managed agent layer

Some services sit between the assistant and your site and add logging, approvals and rollbacks. NibWP is one such option. Others exist. Check how they store credentials, what they log and what happens when you revoke access.

Whichever route you choose, note down two values: the MCP endpoint URL (often a path under /wp-json/) and the username of your dedicated user. You already have the application password.

Step 4: Connect your AI assistant

The exact screens depend on the assistant and change over time, so follow the current instructions from the assistant’s vendor for adding an MCP server or connector. The general shape is the same everywhere.

  1. Open the assistant’s settings and find the place to add an MCP server or custom connector.
  2. Give it a name you will recognise, such as WordPress staging.
  3. Provide the endpoint. For a remote HTTP server that is the URL. For a local proxy it is a small command with environment variables for the site URL, username and application password.
  4. Provide credentials the way the setup asks. Many HTTP setups use a Basic authorization header made from the username and application password. Some use OAuth, in which case you approve access in a browser instead.
  5. Save, then restart or reload the assistant if it asks you to.

The step-by-step for Claude specifically is in how to connect Claude to WordPress through MCP, and the ChatGPT route is in the ChatGPT guide. If your site runs a newer WordPress release, the WordPress 7.0 AI client and connectors explainer describes the other direction, where WordPress itself calls AI providers.

If the connection fails

Work through the likely causes in this order:

  • 401 or 403 errors. The username or application password is wrong, or the user lacks permission for that ability. Recreate the password and check the role.
  • 404 on the endpoint. The MCP route is not active, or pretty permalinks are off. Check the plugin settings and Settings, then Permalinks.
  • Timeouts or blocked requests. A security plugin or firewall may be blocking REST calls. Check its logs before you disable anything.
  • The assistant shows no tools. The abilities may not be marked public for MCP, or the client needs a restart.
  • Local proxy will not start. Confirm Node.js is installed and that the command path in your config is correct.

Step 5: Run a read-only first task

A good first task is one that cannot hurt anything. You want proof that the connection works and a feel for how the agent reasons. Try these in order and watch what the assistant says it is doing.

  1. “List the tools you can use on my WordPress site.” You should see the abilities you exposed. If the list is longer than you expected, stop and look at why.
  2. “Show me the 10 most recent posts with their status and word count.” This confirms read access.
  3. “Which posts have no excerpt or no featured image?” This is a real audit task. It saves time without changing anything.
  4. “List installed plugins and tell me which ones are inactive.” Only works if you exposed plugin abilities and use an admin role on staging.

Cross-check at least one answer against wp-admin. Agents can be confidently wrong, and learning how they fail on easy questions teaches you where to be careful on harder ones. The prompts in our maintenance prompt library are written for exactly this kind of audit work.

Step 6: Try a low-risk write task

Once reads work, let the agent write something you can easily undo. The safest write is a draft.

  1. Ask: “Create a draft post titled ‘Agent test’ with a two-sentence summary of what an application password is. Do not publish it.”
  2. Open wp-admin and find the draft. Check the title, content, author and status.
  3. Ask the agent to change one thing, such as the title, and confirm the revision history shows the change.
  4. Delete the test draft yourself.

Notice who WordPress lists as the author. It should be your dedicated agent user. That is the audit trail working. If your setup writes to an activity log, check that the action is recorded there too.

If you want something more useful, move to a small real job: write meta descriptions for five old posts and save them as suggestions in a draft note, not into the live fields. Reviewing five items takes minutes and tells you whether the output quality is good enough to scale up.

Safety rules for a first agent

An agent is fast, which is the benefit and the risk. A wrong instruction can change a hundred posts before you notice. These rules are cheap and they prevent most problems. For the full threat picture, read the security risks of AI agents on WordPress.

Use staging first, then production with eyes open

Run everything on a copy until you trust the setup. When you move to the live site, follow a workflow where changes are reviewed before they go live. The AI agent staging workflow lays out one way to do that.

Keep permissions as small as the task

Start with Editor or lower. Move to Administrator only for a defined job, such as a plugin audit, and downgrade the user afterwards. Remember that an ability’s permission check applies to the user, so the role is your main control.

Require approval for destructive actions

Deleting posts, changing settings, updating plugins and editing users should never happen without you seeing and approving them. Some tools offer approval prompts for deletions. Turn them on. Where the tool does not, keep the agent user’s role too low to delete.

Back up before every session that writes

This is a habit, not a one-off. A backup taken in the minute before a session is the cheapest insurance you can buy.

Protect the credential

Store the application password in a password manager. Do not paste it into chats, screenshots or public repositories. If you suspect it leaked, revoke it from the user profile. It takes seconds and does not affect your own login.

Keep the agent off the public internet where you can

Use HTTPS, and where your host allows it, restrict who can reach the admin and REST endpoints. Rate limiting and a security plugin help. See the WordPress security guide for the basics that protect any site, with or without an agent.

Read what the agent says it did

Assistants summarise their own actions, and sometimes the summary is generous. Check the result in wp-admin. For content, look at the revision history. For settings, compare before and after.

Tasks that suit a first agent

After the test draft, choose jobs where the cost of a mistake is low and the time saved is real. These tend to work well:

  • Content audits. Find posts missing excerpts, featured images, alt text or internal links. The output is a list, which is easy to review. The guide to refreshing old content shows what to do with the list.
  • Alt text drafts. Generate suggestions for images that lack them, then approve them. See AI image alt text on WordPress.
  • Internal link suggestions. Ask for related posts to link from a given article, then add the best ones yourself. The internal linking guide explains how to judge them.
  • Maintenance reports. Summarise pending updates, inactive plugins and failed scheduled tasks. Pair this with a maintenance checklist so you know what matters.
  • Draft pages from a brief. Create a layout as a draft and refine it yourself. Building a landing page with an AI agent walks through a fuller example.

Tasks that are better left alone at first: editing the checkout, changing payment settings, touching user accounts, bulk-deleting anything and running database changes. If you later need maintenance through the command line, see WP-CLI and AI agents for maintenance, and learn its failure modes before you try it.

MCP, REST API or WP-CLI: which route fits you

You may wonder why MCP at all, when WordPress already has a REST API. The REST API is a general-purpose interface for developers. MCP is a standard way for AI assistants to discover and call tools, and abilities add descriptions and permission checks that machines can read. In practice you often use both: the MCP Adapter’s HTTP transport is built on REST routes. For a comparison, read MCP versus the REST API for WordPress.

RouteGood forWatch out for
MCP over HTTPAssistants that support remote connectors, simple setupCredential handling, what tools you expose
MCP over STDIO with WP-CLILocal or server access, developersShell-level power; keep it to staging at first
Plugin with a built-in MCP serverBeginners who want a settings screenLarge tool lists; read what is exposed
Managed agent serviceLogging, approvals, rollbacksA third party in the loop; check data handling

Common mistakes to avoid

  • Connecting with the main admin account. It works, and it removes your safety net. Create a separate user.
  • Starting on the live site. Your first prompts will be clumsy. Let them be clumsy on staging.
  • Giving broad instructions. “Improve my SEO” invites changes you did not intend. Ask for one defined thing, with a stated limit, such as “propose” or “draft only”.
  • Skipping the review. The agent’s confidence is not evidence. Open the result.
  • Leaving access open. When the experiment ends, revoke the application password or delete the user.
  • Exposing everything. More tools mean more ways to go wrong. Expose what the current task needs.
  • Treating AI text as finished. Drafts need an editor. See how to use AI content without hurting SEO.

Cost, privacy and data: what leaves your site

Whatever the agent reads goes to the AI provider as part of the conversation. If a task reads customer data, order details or private drafts, that content is sent to the provider you chose. Before you connect a live store or membership site, read the provider’s data policy and check whether your plan excludes your data from model training. If you serve visitors in regions with strict privacy law, ask a qualified professional what you must disclose. This guide cannot give legal advice.

On cost, an MCP connection itself is usually free, but the assistant plan or the API usage behind it is not. Some plugins use your own API key and bill by usage. Others include a free allowance. Check the vendor’s current pricing before you run a large batch job, because a loop over thousands of posts can use far more than a single chat.

What to do after your first 20 minutes

You now have a working connection. Here is a sensible path for the next few weeks.

  1. Week one. Run read-only audits on staging. Compare the agent’s findings with your own knowledge of the site.
  2. Week two. Let the agent create drafts and suggestions. Review each one. Track how often you have to fix something.
  3. Week three. Move one narrow task to production with a backup and a defined scope, such as adding missing excerpts to 20 posts.
  4. Ongoing. Review the abilities you expose every month. Remove what you do not use. Rotate or revoke application passwords you no longer need.

If you later need more search visibility from AI tools reading your site rather than the reverse, see AI search visibility for WordPress. It is a different problem, but the same fundamentals of clear structure and accurate content apply.

FAQ

Do I need to know how to code to set up an AI agent on WordPress?

No, for the plugin-based and connector-based routes. You create a user, generate an application password and paste a URL and credentials into your assistant’s settings. Coding helps if you use the official MCP Adapter and want to register your own abilities, since that involves PHP. For a first agent that reads and drafts content, a plugin that ships an MCP server is enough, and the safety steps in this guide apply either way.

Is it safe to give an AI assistant access to my WordPress site?

It can be, if you limit what it can do. Use a dedicated user with a low role, a revocable application password, HTTPS, a staging copy for early tests and backups before write sessions. The risk comes from broad permissions and unreviewed changes, not from the connection itself. No setup is risk-free, so keep approvals on for deletions and settings changes.

What is an application password and is it the same as my login?

No. An application password is a separate credential WordPress generates for one program. It lets that program authenticate as a user without knowing the user’s real password, and you can revoke it without changing the login. It carries the same permissions as the user it belongs to, which is why you should create it for a low-privilege user.

Which AI assistants work with WordPress through MCP?

Any assistant or client that supports MCP can connect in principle. Claude and ChatGPT both have ways to add MCP servers or connectors, and several coding assistants do as well. Availability depends on your plan and workspace settings, and these features change often. Check the vendor’s current documentation, and see our Claude and ChatGPT connection guides for the usual steps.

Can the agent delete my posts or break my site?

Only if its user has permission and the tools are exposed. A Contributor cannot publish or delete others’ posts. An Administrator can do far more. That is why you start with a low role, expose only needed abilities, require approval for destructive actions and keep backups. If something does go wrong, restore from backup and review what the agent was allowed to do.

Should I connect to my live site or a staging copy first?

Staging first. Your first prompts will be imprecise, and you will learn how the agent behaves. A copy lets you make those mistakes safely. Once the workflow is solid, move one narrow, reviewed task to the live site, with a fresh backup. Many hosts and plugins offer one-click staging; see our staging guide for options.

How do I stop an agent from having access?

Revoke the application password in the user’s profile, or delete the dedicated user. Either takes effect immediately for new requests. Also remove the server or connector from the assistant’s settings, and turn off the MCP plugin or the exposed abilities if you no longer need them. Doing all three leaves no stale access behind.

Will an AI agent hurt my SEO?

Not by itself. Quality does the damage or the good. Thin, unedited, mass-produced content can hurt. Careful audits, better meta descriptions, fixed alt text and edited drafts can help. Review everything, add real expertise and avoid bulk publishing. Our article on AI content and SEO goes into the details.

Conclusion and next steps

Your first AI agent on WordPress is mostly a safety exercise with a small amount of configuration. You made a dedicated user, generated an application password, connected an assistant through MCP, ran a read-only task and tried a reversible write. The habits matter more than the tools: staging first, the lowest role that works, backups before writes, review after every change and revocation when you are done.

From here, pick one repetitive job on your site, such as finding posts without excerpts, and let the agent handle only that for a week. If the results hold up, widen the scope slowly. If you want a broader map of the field, start with the complete guide to AI and MCP on WordPress, then read the security and staging articles before you let an agent near production.

This article is coming soon.

Keep reading

Let your AI assistant do this for you

Open the ready prompt with step‑by‑step instructions for running it through NibWP.

← Back to the blog

SponsoredNibWP

Stop clicking. Start telling.

Ask your AI to build pages, fix speed and nail SEO, and NibWP does it on your WordPress site in minutes.

Try NibWP.com

Get the WPMusk Sunday brief

One email a week with our newest tutorials, honest tool reviews and resources you won’t find on the blog. No spam, unsubscribe in one click.

Newsletter (Sunday brief)