AI Agents

• Video lessons
• Connecting an agent
  • Turning it on
  • The address and the key
  • claude.ai
  • On the same machine
• How it works
• What an agent can do
  • Content, media and settings
  • Tools from modules
  • What to ask
• Safety
• Configuration
• Tools of your own
• Next

Connect Claude, Cursor or any other MCP client to the site and it runs the site for you: posts, media, settings, and everything the installed modules add. It works with the rights of the user whose key it holds — nothing more.

The other way: AI inside the admin

Here your agent, with your subscription, works on the site from the outside. For the site itself to ask a model — a "Translate with AI" button, with an API key saved on the site and a monthly budget — see AI Writer. The same tools in a chat inside the admin, with the site's model, are the Assistant; Three Ways to Use AI compares them.

Video lessons

First connect an agent, then watch it at work: clearing out demo content and bringing a post back, filling the site with sample content, SEO and redirects, the day's orders, the shop's sales, the site's health. The chat on the left is a real agent's, and every tool it calls changes the site on the right.

Clear the site with an AI agent

Fill the site with sample content

Orders with an AI agent

Shop sales with an AI agent

SEO and redirects with an AI agent

Site health with an AI agent

Connecting an agent

Turning it on

1. AI agents → Settings: turn on AI agents. Until then /mcp answers 404.

2. The permission Let AI agents work with my application passwords (use_ai_agents). Administrators have it; give it to another role on the roles screen only if that role should have agents.

3. A key — an application password of the REST API, shown once:

   • Filament AI agents → Connect an agent makes one for you, with the settings ready to paste into the agent;
   • Nova AI agents → Connect an agent shows the address and the settings, and links to REST API → Add New Application Password for the key.

   Any application password works the same way.

A key that writes lets the agent create and change records; a read-only key gets the reading tools only.

The address and the key

The address is https://your-site/mcp; the key goes in the Authorization header as Bearer <id>|<secret>.

Claude Code

claude mcp add --transport http kitloom https://your-site/mcp --header "Authorization: Bearer 12|…"

Cursor

// .cursor/mcp.json
{
  "mcpServers": {
    "kitloom": {
      "url": "https://your-site/mcp",
      "headers": { "Authorization": "Bearer 12|…" }
    }
  }
}

Claude Desktop

// claude_desktop_config.json, through mcp-remote
{
  "mcpServers": {
    "kitloom": {
      "command": "npx",
      "args": ["mcp-remote", "https://your-site/mcp", "--header", "Authorization: Bearer 12|…"]
    }
  }
}

claude.ai

claude.ai — on the web, on phones and as Claude Desktop's own connectors — takes the key in a header, like the other clients. In Settings → Connectors → Add custom connector:

Field	What goes in
Name	anything, e.g. My site
URL	https://your-site/mcp
Authentication	No sign-in
Request headers → Add header	Authorization, with the value Bearer 12|…

Then, in a new chat, turn the connector on in the tools menu and ask away. Connect an agent — and the dashboard, for a key made ahead — shows the URL and the header's value ready made: in Filament with a new key, in Nova with a place for yours.

A client without headers

A client that cannot send a header can take the key inside the address instead, with the bar of the key written as a dot:

https://your-site/mcp/k/12.…

Such an address is a password

It lands in the web server's logs, in browser histories and in error reports that keep the URL, so use the header wherever the client has one. Make a key of its own for it, with an expiry, and revoke it when it is no longer used: a revoked or expired address answers 404, as one that never was. kitloom-ai.mcp.secret_address set to false turns these addresses off. Signing in with OAuth is not offered yet.

On the same machine

php artisan mcp:start kitloom serves the same tools over stdio — for Claude Code working in the project itself:

claude mcp add kitloom -e KITLOOM_AI_USER=admin -- php artisan mcp:start kitloom

It acts as the user KITLOOM_AI_USER names (login or email), who needs use_ai_agents too; with nobody named, every tool refuses. It does not depend on the settings switch: whoever runs it has the console anyway.

How it works

A kitloom site can be run by an AI agent — Claude, Cursor or any other client of the Model Context Protocol. kitloom/ai is an MCP server at /mcp: its tools read and write posts, pages, terms, media and settings, tell what lives at an address and show a page as the site renders it. Installed modules add tools of their own — menus, widgets, SEO, redirects, custom fields, forms, comments, the shop, sliders, bookings, courses, listings, languages, webhooks, maintenance mode, backups, the activity log, Site Health and more.

The agent works with the rights of the user whose key it holds, through the same code the admins use. It cannot do what that user may not do, it cannot delete anything for good, and a revoked key stops it at once.

Package	Gives
kitloom/ai	The MCP server, the core tools, the local server for the console
kitloom/ai-filament / kitloom/ai-nova	The AI agents section: Connect an agent, Settings, How it works

composer require kitloom/ai kitloom/ai-filament   # or kitloom/ai-nova

It needs the REST API module: its application passwords are the agents' keys.

What an agent can do

Content, media and settings

The core tools, always there:

Tool	What it does
describe_site	The site, its post types and taxonomies with what they support, the statuses, what the key may do with each type. The agent's first call
list_content	Records of a post type, every status but the trash: search, status, terms, parent, author, dates, pages
get_content	One record in full, with its stored HTML
create_content	A new record — a draft unless the agent gives a status — with terms, a featured image and, for types with comments, whether they are open
update_content	Changes to a record: only the fields sent
trash_content, restore_content	Into the trash and back, with the status it had
content_revisions, restore_revision	A record's saved versions — when, by whom, whether an agent made it, what changed — and one brought back; the restore is a version too, so it can be undone the same way
list_terms	Terms of a taxonomy: search, parent, the terms of a record; or those in the trash
create_term, update_term	A new term, changes to one
trash_term, restore_term	A term into the trash and back — never the default category
list_media	The media library with the address of each file; images without alt text, and files nothing seems to use
upload_media	A file into the library — from a public web address, or sent as base64 with its name — with title, alt text and caption
update_media	Title, alt text or caption of a file
list_users	Users by name, for author ids — without emails
list_settings	The declared settings by group, with values and whether the agent may change each
update_setting	One setting, checked against its declaration and the user's permission
resolve_url	What lives at an address: the kind of page, the record or term, the records of an archive, the templates the theme tries and the one it has
render_page	A page as the site shows it — the text a visitor reads, the HTML, or the page's parts as JSON (an object, page); a draft as its preview
get_page_schema	The JSON Schema of that page, or of one part — the parts this site has installed, the theme's own among them

Tools from modules

An installed module adds its tools; an agent sees them only when its user has the module's permission.

Module	Tools	What for
Navigation	menus_list, menu_get, menu_save	Menus and their places
Widgets	widgets_list, widgets_save	The widgets of each zone; a widget left out goes to the inactive ones
SEO	seo_get, seo_update	Title, description, robots, canonical, Facebook card, primary category — of a record or a term
Redirects	redirects_list, redirects_create, redirects_update, redirects_trash, redirects_settings_get, redirects_settings_update, not_found_list, redirect_from_404	Redirects, their settings and the 404 log
Custom Fields	fields_get, fields_set, field_groups_get, field_groups_save	Field values, checked as the edit form checks them; field groups made and fields added or changed — never removed
Forms	forms_list, form_get, form_create, form_update, submissions_list, submission_spam	Contact forms and their messages; a new form mails the site's admin — where messages go is a person's choice
Comments	comments_list, comments_moderate, comments_reply	Comments and product reviews: list, approve, hold, spam, trash, restore, answer
Shop	products_list, product_get, product_save, product_category_image, attributes_list, attribute_save, orders_list, order_get, order_create, order_set_status, order_add_note, sales_report, top_products, customers_list, coupons_list, coupon_save	The catalogue, its attributes, the orders, the sales and the customers, the coupons. A new product is a draft; an order taken by phone starts pending or on hold at the catalogue's prices — nothing is charged; a completed or refunded order is not reopened; refunds stay in the admin; notes are internal
Sliders	sliders_list, slider_get, slider_save	Sliders and their slides
Bookings	bookings_catalog, bookings_slots, bookings_list, bookings_create, bookings_decide, bookings_clients, bookings_client, bookings_client_note, bookings_service_save, bookings_staff_get, bookings_staff_save, bookings_close	Services and free times, appointments and their decisions, the clients and notes about them; the services, the staff with their hours and time off, the salon closed for a holiday. Nothing is deleted, and which account a staff member signs in with stays a person's choice
Backup	backup_status, backup_run	The backups and a new one; archives are downloaded and deleted in the admin
Activity Log	activity_list	Who changed what
Site Health	health_report, health_run, health_failed_jobs	The checks and their results; the queue's failed jobs — which and when, never the message
Languages	languages_list, languages_get_translations, languages_save_translation, languages_translate	The site's languages and the translations of posts, terms and users; a machine translation is saved as a draft
Link Checks	link_checks_list, link_checks_recheck, link_checks_dismiss, link_checks_replace	Broken links with the posts they are in: check again, dismiss, replace or unlink an address everywhere
Post Types	post_types_list, post_types_get, post_types_create, post_types_update, post_types_trash, post_types_restore	Post types and taxonomies made in the admin, checked as the form checks them; a trashed type keeps its records
Search	search_status, search_reindex	How the search index is doing, and a rebuild on the queue; its settings are options of update_setting
Image compression	media_optimizer_status, media_regenerate	Which compression tools the server has and how to install the missing ones; the sizes of a few images cut — and compressed — again, no file deleted
Import & Export	imports_list, import_get, import_start	The history of imports, and an import of a WXR file from a public address; exports stay in the admin
Courses	courses_list, courses_get, courses_save, courses_enroll, courses_students	Courses with their curriculum, enrolling a user by hand, the students and their progress. A new course is a draft; lessons and quizzes are written with the core tools
Listings	listings_search, listings_get, listings_save, listings_enquiries, listings_enquiry_status	Properties and directory listings, and the enquiries sent from their pages. A new one is a draft
Webhooks	webhooks_list, webhooks_create, webhooks_update, webhooks_deliveries	Outgoing webhooks, a test delivery and the delivery log. The secret is shown once, on creation; a webhook is disabled, not deleted
Maintenance mode	maintenance_settings_get, maintenance_settings_update	The holding page — coming soon or maintenance — its words, the roles that still see the site, the access link
Users (core)	users_create, users_update, users_set_role	A user invited by email to choose their own password, a public profile, a role — never an address or a password, never a role that runs the site or manages users, nor anyone who holds one
Search & Replace	search_replace_dry_run, search_replace_run, search_replace_runs	A change counted first, then made exactly as counted — in posts, comments, terms and their meta only; options and user meta stay with the admin
Internal Links	internal_links_report, internal_links_list, internal_links_recount	Link counts, the orphans nobody links to, the links of a post and to it
Privacy	privacy_requests_list, privacy_request_create, privacy_request_complete	Personal data requests; a new one is always confirmed by the person; exports and erasures stay in the admin
AI Writer	ai_writer_usage	The month's spending against the budget — never the key
Pulse	pulse_report	Slow pages, queries and jobs, and the exceptions thrown
Payments	payments_methods	Which ways to pay are on and which credentials are missing — never a value
Site (core)	site_cache_clear	The page cache cleared, as Settings → Cache's button
Site (core)	theme_customize	The active theme's settings — colours, the logo, the footer's line — as on Appearance → Customize, published at once; describe_site lists them under theme_settings

describe_site also lists the shortcodes the site draws, each with an example, so an agent writing content puts in one the site knows.

describe_site also names each module's tools and what the module tells about the site: which field groups show where, the menu places, what each SEO name means. Each module turns its own tools off with kitloom-<name>.ai.register.

What to ask

A few things an agent does well on a kitloom site:

• "Write a draft post about our spring collection, put it in News, and take the cover from this link."
• "Find the pages that still mention the old phone number and fix them."
• "Which addresses gave 404 this week? Redirect the ones that have a new page."
• "Fill in the SEO descriptions of the products that have none."
• "Show me how the About page looks now, then shorten its first section."

Safety

Who sees and writes what. An agent sees with the eyes of the key's user, type by type: every record of a type whose others' records the user may edit (edit_others_posts, edit_others_pages, …), otherwise what is published and the user's own. Only the declared post types are read through the content tools — a module's own records come through its tools and permission — and the text of a password-protected record only for a user who may edit it. It writes as the key's user, through the same writers the admins use: the slug is made unique, the dates are filled, and whatever the user may not do — publish, edit another author's record, manage terms — is refused with a message the agent reads. The activity log shows an agent's changes under the key's user.

Nothing is deleted for good. The agent moves records to the trash; permanent deletion is left to a person in the admin.

Every change can be undone. Each save of a record keeps a revision, as WordPress keeps them — the title, the content, the excerpt, the custom fields and the SEO meta. A revision an agent's change left is marked so: content_revisions lists them, restore_revision and the Revisions panel of both admins bring an older version back, and the restore is a revision too. A search and replace keeps each record as it was before the run.

Settings agents leave alone. Protected settings — the site address, the roles, secrets — are never shown to an agent. Some more are locked: the agent sees them without their values and may not change them, because text it reads could try to talk it into it. These are who gets in (registration, the default role, the admin email, two-factor, social login), the doors agents and apps come through (AI agents, REST API), and the code on every page and the data that leaves the site (Header & Footer, Tag Manager, Analytics, Ad Pixels, Sentry, Backups, the mail server). A person changes those in the admin; kitloom-ai.settings.locked holds the list.

Files from the web. upload_media fetches only from public hosts, redirects included, and — from an address or sent as base64 — takes only file types the library takes, up to its size limit (kitloom-media.max_size, kitloom-media.remote).

Every request carries its key. The address has no session and no CSRF. Requests are limited per user: kitloom-ai.mcp.rate_limit a minute (120; 0 — no limit).

Configuration

php artisan vendor:publish --tag=kitloom-ai-config

Key	Default	Meaning
mcp.register	true	The MCP server at all
mcp.path	mcp	Its address
mcp.rate_limit	120	Requests a minute per user; 0 — no limit
mcp.secret_address	true	The /mcp/k/… addresses for claude.ai
local.user	KITLOOM_AI_USER	Who the local server acts as
render.max_chars, render.max_chars_limit	20000, 200000	The default and the largest length of render_page
settings.locked	the list above	The settings agents may not change; * matches any end
filament.register, nova.register	true	The admin sections
filament.dashboard, nova.dashboard	true	Connect an AI agent at the top of the dashboard, while agents are on, for a user who may use them and has no key yet. Nova cannot add to the host's dashboard: there the host lists Kitloom\Ai\Nova\Cards\AgentCard in its Main, and false hides it

Tools of your own

A module offers tools without requiring kitloom/ai: a Kitloom\WpSchema\Contracts\AgentTool in its domain package, tagged AgentTool::TAG, and what the agent should know first as an AgentContext. See Extension Points → AI agents and Testing → Agent tools.

Next

• Headless & APIs — page JSON and the REST root
• Module Catalog