Ultra Provider

A WordPress plugin that lets an administrator register any OpenAI-compatible API endpoint as a provider for the WordPress AI Client, pull its model list, and choose which models the site's AI features may use.

The other provider plugins in this family ship one hard-coded provider each. Ultra Provider makes the provider registry user-extensible.

What it does

• Registers user-defined providers — every enabled entry becomes a real AI Client provider, so it appears on the Connectors screen, in the AI plugin's model pickers, in the /ai/v1/providers REST route and in Bestony's AI Settings.
• Delegates credentials to the Connectors API — the API key lives in the standard connectors_ai_{id}_api_key option, with the usual masking, validation and {ID}_API_KEY constant/environment override. The provider is only "connected" while the endpoint actually answers.
• Pulls and curates models — GET /models is fetched from the admin screen; the administrator ticks models and labels each as text, text + vision, or image generation. Manual entry is available for endpoints that hide models.
• Applies user-chosen defaults — the selected text, vision and image models are moved to the front of the wpai_preferred_*_models lists.
• Handles per-endpoint quirks — structured output mode, token parameter name, extra headers (with a {key} placeholder), request timeouts, and key-less local endpoints.

Requirements

• WordPress 7.0 or later (the bundled AI Client and Connectors API).
• PHP 7.4 or later.

Install

Copy this directory to wp-content/plugins/ultra-provider/ and activate the plugin. The Ultra Provider menu appears in the admin sidebar.

Usage

1. Add provider — name, provider ID (the AI Client provider ID and connector ID; derived from the name when left empty), API base URL such as https://api.example.com/v1, optional description, credentials URL, authentication mode, structured output mode, token parameter name and extra headers.
2. Save the API key on Settings > Connectors. The key field, its constant name and its environment variable follow core's connector conventions automatically. Key-less endpoints (Ollama, LM Studio) use authentication mode "None" and skip this step.
3. Fetch models on the provider's Models screen, tick the models to expose, and pick each model's capability:
   • Text generation — chat, history, tools and JSON output.
   • Text generation + vision — same, plus image input.
   • Image generation — the OpenAI images endpoint.
   • A model that is not usable can be left unticked.
4. Optional defaults — set a default text, vision and image model so AI features prefer them.

Models can be filtered by capability in every picker: an image feature is never offered a chat model, a vision feature only sees models declared with image input.

How it works

ultra-provider.php            bootstrap: init 5 registration, init 25 auth backfill, preferences, admin
src/
├── Admin/SettingsPage.php    list / add-edit / models / defaults screens, admin-post handlers
├── Config/Repository.php     option store: validation, slots, preferences, header parsing
├── Runtime/
│   ├── Slots.php             20 tiny provider classes (ProviderSlot01 … ProviderSlot20)
│   ├── AbstractDynamicProvider.php
│   ├── CuratedModelMetadataDirectory.php
│   ├── DynamicProviderAvailability.php
│   ├── ProviderAuthentication.php
│   ├── RequestTrait.php
│   ├── DynamicTextGenerationModel.php
│   └── DynamicImageGenerationModel.php
└── Util/                     Keys, Http, AllowedHosts, ModelCatalog, Preferences

Dynamic classes. The AI Client identifies a provider by its class, and a class's metadata() is static, so one class is needed per provider. The plugin ships twenty tiny slots; each slot class carries only its slot number and reads the configuration assigned to that slot. No eval(), no generated files, and a class is loaded only while its slot is in use.

Registration order. Providers register on init priority 5; WordPress core builds Connectors cards at priority 15 from the registry, so the entries appear automatically. At priority 25 the plugin makes sure every registered provider has an authentication instance — an empty one for key-less endpoints, because the OpenAI-compatible SDK base classes always call getRequestAuthentication().

Curated catalog. CuratedModelMetadataDirectory reads the selected models from the option store on every call: no network request, no 24-hour metadata cache, and edits are visible on the next request.

Availability. DynamicProviderAvailability verifies a provider with GET /models (one-token chat completion fallback), cached per slot and credential — a success for an hour, a failure for five minutes. This drives the Connectors "connected" state, the key validation when an admin saves a key, and the model-preference checks.

Safe requests. The AI Client sends everything through wp_safe_remote_request(), which blocks loopback/private hosts and non-standard ports. Util\AllowedHosts opts in exactly the hosts and ports of enabled providers; no other private address is ever allowed.

Options

Per-provider credentials are stored by the Connectors API, not by this plugin.

Development

No build step and no dependencies; the PSR-4 autoloader is a 20-line file.

php scripts/selfcheck.php
php scripts/selfcheck.php --sdk=/path/to/wp-includes/php-ai-client

The self-check stubs the WordPress functions it needs and covers the option store (slots, validation, headers, preferences), the model catalog and request construction (structured output modes, token parameter, custom {key} headers, image response shapes), the availability probe, the safe-request allowances and the preference filters. It exits non-zero on the first failure count.

Limitations

• OpenAI-compatible protocol only. Anthropic-style messages APIs and Google's native API are not supported.
• Twenty providers per site; remove one to free its slot.
• Chat and image models are separate. One model ID cannot be both, matching the AI Client's capability model.
• A provider ID that collides with another plugin's provider or connector is refused; the other plugin wins.

License

GPL-2.0-or-later.