AI Provider for Command Code

AI Provider for Command Code for the WordPress AI Client (WP 7.0+).

Registers Command Code as a first-class AI provider in the WordPress 7.0 Connectors system (Settings → Connectors). No plugin code touches API keys — WordPress derives everything from the provider ID:

Requirements

• WordPress 7.0+ (AI Client and Connectors API are in core since 7.0)
• PHP 7.4+
• A Command Code plan with API access (Provider, Pro, Max, Team, or GOAT)

Installation

Drop the folder into wp-content/plugins/ and activate. The plugin registers with the AI Client at init, and core auto-creates the "Command Code" connector card.

Configuration

Set the key via environment variable or PHP constant (recommended for production):

define( 'COMMANDCODE_API_KEY', 'cmd-...' ); // wp-config.php

…or enter it on Settings → Connectors, where it is stored in the database (masked in the UI; core stores keys unencrypted — see core trac #64789).

Resolution priority: environment variable → PHP constant → database.

Usage

Any plugin on the site can now generate text via the standard API:

$text = wp_ai_client_prompt( 'Summarize the benefits of caching in WordPress.' )
    ->generate_text();

if ( is_wp_error( $text ) ) {
    // Handle error.
}

Limiting the model catalog to your preferences

By default the connector exposes every model Command Code serves on the chat wire (currently ~59). To restrict the catalog to the models you actually want to use — which also narrows automatic selection and any model picker built on the registry — filter ai_provider_for_commandcode_models (e.g. in an mu-plugin or theme functions.php):

// Allowlist: only these models are ever offered or auto-selected.
add_filter(
    'ai_provider_for_commandcode_models',
    static function ( array $models ): array {
        $allowed = array(
            'deepseek/deepseek-v4-flash',
            'MiniMaxAI/MiniMax-M3',
            'gpt-5.5',
        );
        return array_values( array_filter(
            $models,
            static function ( $model ) use ( $allowed ): bool {
                return in_array( $model->getId(), $allowed, true );
            }
        ) );
    }
);

// Or just drop one model you never want to surface:
add_filter(
    'ai_provider_for_commandcode_models',
    static function ( array $models ): array {
        return array_values( array_filter(
            $models,
            static function ( $model ): bool {
                return 'gpt-5.3-codex' !== $model->getId();
            }
        ) );
    }
);

The filter must return a list of the same model objects ($model->getId() gives the model ID). No core or SDK filter exists for this — it is provided by this plugin (see CommandCodeModelMetadataDirectory::applyModelFilters()).

Recipe: one text model + one vision model

To run a single site on exactly one text model and one vision model — e.g. deepseek/deepseek-v4-flash for text and MiniMaxAI/MiniMax-M3 for vision — drop a file like this into wp-content/mu-plugins/:

<?php
/**
 * Plugin Name: Command Code Model Allowlist
 * Description: Restricts the Command Code connector to deepseek/deepseek-v4-flash (text) and MiniMaxAI/MiniMax-M3 (vision).
 */

use WordPress\AiClient\Providers\Models\DTO\ModelMetadata;

add_filter(
    'ai_provider_for_commandcode_models',
    static function ( array $models ): array {
        $allowed = array(
            'deepseek/deepseek-v4-flash',
            'MiniMaxAI/MiniMax-M3',
        );
        return array_values( array_filter(
            $models,
            static function ( ModelMetadata $model ) use ( $allowed ): bool {
                return in_array( $model->getId(), $allowed, true );
            }
        ) );
    }
);

The vision model stays available for image-description requests because isVisionCapable() still marks MiniMaxAI/MiniMax-M3 as vision-capable — the allowlist only removes entries, it doesn't change their metadata. Text-only requests will never select it since input-modality matching still applies; it's simply the only vision-capable model left in the catalog.

Cache gotcha: the model list is cached for 24 hours in the wp_ai_client object cache group (WP_AI_Client_Cache, backed by whatever persistent object cache the site uses — e.g. Redis). Adding, removing, or editing this filter has no visible effect until that cache entry expires or is cleared. Flush it immediately after deploying a filter change:

wp_cache_flush_group( 'wp_ai_client' );

(or wp cache flush if the site's object cache backend doesn't support group-only flushing).

Picking a model per request

// Preference list: first available model wins; falls back to any compatible model.
$text = wp_ai_client_prompt( 'Refactor this PHP snippet.' )
    ->using_model_preference( 'deepseek/deepseek-v4-flash', 'gpt-5.5' )
    ->using_temperature( 0.2 )
    ->generate_text();

Vision: describing images, not generating them

"Vision" here means a model that takes an image as input and returns a text description or analysis (e.g. alt text). Command Code's API has no image-generation models, and this provider never advertises image generation.

The default vision model is MiniMaxAI/MiniMax-M3 (verified vision-capable on the Command Code chat wire). Image-description requests resolve to it automatically, since text-only models are filtered out of vision requests:

$alt = wp_ai_client_prompt()
    ->with_text( 'Write concise alt text for this image.' )
    ->with_file( $image_file, 'image/png' )
    ->generate_text();

To prefer specific vision models, list them explicitly — vision requests only consider models verified to accept image input:

$alt = wp_ai_client_prompt()
    ->with_text( 'Write concise alt text for this image.' )
    ->with_file( $image_file, 'image/png' )
    ->using_model_preference(
        'MiniMaxAI/MiniMax-M3',
        'google/gemini-3.8-flash',
        'gpt-5.4'
    )
    ->generate_text();

Model catalog & classification

Model metadata is fetched live from GET /provider/v1/models and classified in code (src/Metadata/CommandCodeModelMetadataDirectory.php). Every classification rule was verified against the live API on 2026-09-08:

• *Anthropic wire (`claude-) is excluded** — those models require Command Code's/provider/v1/messages` endpoint, which v1 of this provider does not implement (planned).
• Text generation + chat history are advertised for every chat-wire model.
• temperature / top_p: accepted by every family tested (DeepSeek, Qwen, Gemini, Kimi, GLM, MiniMax, Grok, GPT-5.x, and more).
• presence_penalty / frequency_penalty: advertised everywhere except the Google Gemini family (the only tested family that rejects them).
• Vision (image input) is whitelisted per model — currently: deepseek/deepseek-v4-flash-vision-exp, Google Gemini 3.x, Qwen/Qwen3.8-27B, GPT-5.x chat models, Moonshot Kimi K2.5–K3, MiniMaxAI/MiniMax-M3, Thinking Machines Inkling, z-ai/glm-5.3-flash.
• New or unknown models default to conservative text-only metadata until verified, so the connector keeps working as Command Code's catalog churns.

Zero data retention

Send x-cmd-zdr: 1 on every request by attaching request options:

use WordPress\AiClient\Providers\Http\DTO\RequestOptions;

$text = wp_ai_client_prompt( '…' )
    ->using_request_options( new RequestOptions( headers: [ 'x-cmd-zdr' => '1' ] ) )
    ->generate_text();

(Constructor signature may vary by SDK version — see the PHP AI Client docs.)

Development

composer install
composer lint    # phpcs + phpstan
composer test    # phpunit

The fixture tests/fixtures/models-response.json is a live capture of GET /provider/v1/models (67 models, 2026-09-08); refresh it to re-validate the classifier against the current catalog.

License

GPL-2.0-or-later. Not affiliated with or endorsed by Command Code. The Command Code logo is used with permission to identify the service, per Command Code's brand guidelines; it remains the property of its owner.