For developers

How Gatehouse works#

Gatehouse does not wrap or replace AI provider plugins. It uses hooks from the WordPress 7.0 AI Client and the standard HTTP API.

Calls through the WordPress AI Client:

Direct calls (a plugin calling a provider with its own key through the WordPress HTTP API, with no AI Client generation in flight):

When a plugin reads a streamed answer itself (for example with its own CURLOPT_WRITEFUNCTION), the body WordPress hands back is empty, so the call is logged without tokens or cost, with a note. Calls made with other HTTP libraries, or from the browser, aren't visible.

Notes:

• Blocked is_supported_*() checks are not logged. When a source is blocked, wp_ai_client_prompt()->is_supported_*() returns false silently, so plugins can hide their AI features without filling the log.
• Where the code lives:
  • includes/class-gateway.php: the hooks;
  • includes/class-redactor.php: personal data detection and redaction;
  • includes/class-provider-adapters.php: provider request formats;
  • includes/class-ledger.php: storage and monthly totals;
  • includes/class-stats.php: dashboard numbers.

Filters and actions#

gatehouse_provider_hosts#

Maps API hosts to provider IDs. Requests to these hosts are checked for personal data, and text-generation requests made directly by plugins are recorded and can be blocked. Defaults: api.anthropic.com, api.openai.com, generativelanguage.googleapis.com, openrouter.ai, api.x.ai, api.mistral.ai, api.deepseek.com, api.groq.com, api.perplexity.ai.

add_filter( 'gatehouse_provider_hosts', function ( $hosts ) {
	$hosts['llm.internal.example.com'] = 'self-hosted';
	return $hosts;
} );

Adding a provider#

Add its host with gatehouse_provider_hosts, as above.

• Detection and redaction apply to every POST request to that host, except model listings (paths ending in /models).
• Direct calls are recorded, and can be blocked, when the URL path matches one of these text generation formats:

gatehouse_price_source_url#

The URL of the price list used by automatic price updates. It defaults to OpenRouter's public model list. A replacement must return the same format: { "data": [ { "id": "provider/model", "pricing": { "prompt": "<USD per token>", "completion": "<USD per token>" } } ] }.

gatehouse_default_prices#

Changes the built-in price table: USD per million tokens, as [input, output]. Prices saved in Settings → Model prices still take priority.

add_filter( 'gatehouse_default_prices', function ( $prices ) {
	$prices['my-fine-tuned-model'] = array( 3.0, 12.0 );
	return $prices;
} );

gatehouse_attribution_skip#

Plugin slugs to skip when working out who made a call. Use it for your own "middleman" plugins that make AI calls on behalf of other plugins, so the call is attributed to the plugin further up the call stack.

add_filter( 'gatehouse_attribution_skip', function ( $slugs ) {
	$slugs[] = 'my-ai-helper-library';
	return $slugs;
} );

gatehouse_record_row (filter)#

Adjusts a call's row just before it is saved, for example to mark a call answered from your own cache: cached = 1, the cost moved into saved, and cost = 0. Cached calls show as "From cache" in the dashboard.

gatehouse_local_providers (filter)#

Provider IDs whose calls cost nothing because the model runs locally. Defaults to webllm.

gatehouse_admin_enqueue (action) and window.gatehouse#

Fires after the dashboard script is enqueued. Pass the handle (gatehouse-admin) as a dependency of your own script.

• Shared components: your script can use the dashboard's components from window.gatehouse: ui, charts, Icon, format, hooks, colors and settings.
• Adding a page: register it with the gatehouse.routes filter from @wordpress/hooks:

import { addFilter } from '@wordpress/hooks';
addFilter( 'gatehouse.routes', 'my-addon', ( routes ) => [
	...routes,
	{ id: 'my-page', label: 'My page', icon: 'spark', Page: MyPage },
] );

gatehouse_pro_active (filter)#

Tells the dashboard that an add-on providing response caching is active.

gatehouse_connector_approval_active, gatehouse_docs_url, gatehouse_support_url (filters)#

• gatehouse_connector_approval_active overrides the detection of the AI plugin's Connector Approval feature.
• gatehouse_docs_url sets the "Full documentation" link on the Help page.
• gatehouse_support_url sets the "Get support" link on the Help page.

Demo mode#

Demo mode is per user (user meta gatehouse_demo_mode) and only applies inside Gatehouse's own REST requests. While it applies:

• Gatehouse_Ledger::table() returns {prefix}gatehouse_demo_requests;
• settings come from gatehouse_demo_settings;
• month totals use gatehouse_demo_spend_* options.

Run code against the sandbox with Gatehouse_Demo::run( $callback ).

gatehouse_recorded (action)#

Fires after each call is logged, with the row ID and the row's values. Use it to forward usage to your own monitoring.

add_action( 'gatehouse_recorded', function ( $id, $row ) {
	if ( 'error' === $row['status'] ) {
		error_log( "AI call failed for {$row['source']}: {$row['note']}" );
	}
}, 10, 2 );

$row keys:

• created_at, source, user_id, provider, model, capability, status (ok, blocked or error);
• input_tokens, output_tokens, cost, priced, latency_ms;
• channel (ai_client or direct), pii (kinds and counts found, for example email:2,phone:1), redactions (items replaced);
• note, prompt_excerpt, response_excerpt.

Source IDs#

Sources are identified as type:slug:

WP-CLI#

REST API#

The dashboard uses these routes under /wp-json/gatehouse/v1. All of them require the manage_options capability.

Building from source#

Requires Node.js. PHP and Docker are not needed: development uses WordPress Playground.

npm install
npm run build        # or: npm start (watch mode)
npm run lint:js

Local site with sample data:

npx @wp-playground/cli@latest server --port=9400 \
  --mount=.:/wordpress/wp-content/plugins/gatehouse \
  --blueprint=dev/blueprint.json --login

Integration tests. tests/run.sh runs them in a throwaway WordPress, with the real Anthropic provider plugin talking to a mock API. Note: parts of the suite still describe the 1.x behaviour (brand brief, redaction on by default) and need updating before they pass.

Documentation screenshots. Run with the local site above running; uses your installed Chrome:

node dev/screenshots.mjs

Release zip:

bin/build-zip.sh     # writes dist/gatehouse.zip

Release checklist#

1. Review the built-in model prices against each provider's official price page. They're the fallback when automatic updates are off or a model isn't in the live list:
   • update the table in includes/class-pricing.php;
   • set Gatehouse_Pricing::CHECKED to today's date;
   • add any new models.
2. Bump the version in gatehouse.php (header and GATEHOUSE_VERSION) and in the Stable tag line of readme.txt.
3. Add a changelog entry in readme.txt. Mention the price review, for example "Model prices checked 2026-12-01; added gpt-5.6".
4. Run npm run lint:js and the Plugin Check plugin against the zip (and tests/run.sh once it's updated).
5. Run bin/build-zip.sh, and refresh the screenshots if the interface changed.
6. Publish a GitHub release with the zip attached, so the website's download button (which links to the latest release) serves it.

Sites with automatic price updates stay current on their own. For everyone else, plan a release every few months that refreshes the built-in table. The dashboard warns users when built-in prices are more than 120 days old.