Claude, Gemini and OpenAI in WordPress: AI Connectors, AI Client and MCP Explained

Originally published March 5, 2026 — Updated October 14, 2026

WordPress now has an official, provider-agnostic AI architecture. WordPress 7 includes the AI Client and Connectors API; separate connector plugins link a site to Claude, Gemini, OpenAI or other model providers; the canonical AI plugin adds ready-to-use features in Gutenberg and the administration interface; and the MCP Adapter lets external AI agents discover and invoke approved WordPress abilities.[^1][^2][^3]

These components are complementary, but they do not serve the same purpose. Installing an OpenAI, Anthropic or Google connector does not automatically add a ChatGPT-like assistant, allow Claude to administer a website or enable an agent to modify WooCommerce products. It supplies the standardized connection that another plugin, feature or authorized agent can use.

What changed since March 2026

When the first provider plugins appeared, the main story was the PHP AI Client SDK: developers could call different AI providers through one interface. The architecture has since matured significantly.

• WordPress 7.0 integrated the AI Client into Core. Plugins can send prompts through the provider-independent wp_ai_client_prompt() API instead of implementing a separate OpenAI, Anthropic or Google integration.[^4][^1]
• WordPress 7.0 added the Connectors API and the Settings → Connectors screen. It provides a common interface for installing, configuring and authenticating AI provider connections.[^2][^5]
• The canonical AI plugin has become the user-facing experimentation layer. Version 1.4 includes writing, translation, image, SEO, comment-moderation, Markdown and embedding-related features.[^6][^7]
• The Abilities API provides machine-readable WordPress capabilities. Each ability can define its inputs, outputs, execution callback and permission rules.[^8]
• The MCP Adapter is now publicly distributed as a canonical WordPress plugin. It translates publicly exposed abilities into MCP tools, resources and prompts for compatible external agents.[^3]

The result is no longer simply "three plugins connecting three AI providers." WordPress now has distinct layers for model access, user-facing AI functionality and controlled agent access.

The WordPress AI stack

A simple way to understand the architecture is:

• AI Client and connectors bring AI into WordPress.
• Abilities describe what WordPress can do.
• MCP makes approved abilities available to external agents.
• The AI plugin supplies ready-made features for users and acts as a reference implementation for developers.

AI Client in WordPress 7

WordPress 7.0 includes a provider-agnostic PHP API for generative AI. A developer starts with wp_ai_client_prompt(), configures the prompt and required capabilities, then requests the appropriate result. The Core wrapper follows WordPress conventions, returns WP_Error objects and integrates with WordPress HTTP transport, hooks, connectors, settings and abilities.[^1][^4]

$result = wp_ai_client_prompt()
    ->with_text( 'Summarize this article in three concise points.' )
    ->generate_text_result();

if ( is_wp_error( $result ) ) {
    // Handle the error.
}

The API can request supported output types through methods such as text generation, image generation and text-to-speech conversion. Actual availability depends on the installed connector and the selected model's capabilities.[^1]

Why it matters

Previously, a plugin author often had to choose one model provider and maintain a bespoke integration. With the AI Client, a content, SEO, translation or commerce plugin can describe the capability it needs while allowing the site owner to choose the provider.

This reduces provider lock-in at the application layer. It does not make all models identical: prices, context windows, output quality, latency, privacy policies and supported modalities still vary between providers.

WordPress 6.9 compatibility

The provider connector pages still support WordPress 6.9 when the separate wordpress/php-ai-client package is installed. On WordPress 7.0 and later, the underlying AI Client infrastructure is supplied by Core, so developers no longer need to bundle that dependency merely for the PHP API.[^4]

For a new production implementation, WordPress 7.0 or later is the clearer baseline. The canonical AI plugin 1.4 requires WordPress 7.0.3 or later.[^7]

Connectors and credentials

WordPress 7 introduced Settings → Connectors, a centralized interface for AI connections. Anthropic, Google and OpenAI are featured providers, but their implementations remain separate plugins so that they can evolve faster than WordPress Core.[^5][^1]

WordPress Core does not silently send content to an AI provider. A connector must be installed and configured, and WordPress code or an enabled feature must explicitly make the request.[^9]

The normal setup is:

1. Install WordPress 7.0.3 or later. 2. Install the provider connector required by the organization. 3. Open Settings → Connectors. 4. Add the provider API key or configure an approved endpoint. 5. Install the canonical AI plugin if user-facing Gutenberg and admin features are required. 6. Open Settings → AI and enable only the necessary experiments. 7. Test the workflow on staging before enabling it on a production site.

The canonical AI plugin also offers Connector Approvals, which can require explicit administrator approval before a plugin or theme uses a configured connector. Its optional key-encryption experiment encrypts provider keys at rest with libsodium; WordPress notes that disabling the experiment or deactivating the AI plugin restores the keys to plaintext storage.[^6]

For higher-security environments, secrets can instead be supplied through environment variables or constants where supported, kept out of the database and excluded from repositories and backups.

Anthropic connector

The AI Provider for Anthropic connects the WordPress AI Client to Claude. Version 1.0.5 supports text generation, multimodal inputs, function calling, extended thinking and dynamic model discovery.

Current capabilities

• Claude text generation
• Multimodal input on compatible models
• Function calling
• Extended-thinking support
• Dynamic discovery of available Claude models
• Multi-turn preservation of extended-thinking signatures
• Continuation of responses that stop with Anthropic's pause_turn
• Explicit token-limit errors instead of silently returning truncated output

Version 1.0.5 was released on October 1, 2026. It corrected multi-turn extended-thinking behavior, paused-turn continuation and thought-token accounting.

Requirements

• WordPress 6.9 with the separate PHP AI Client package, or WordPress 7.0+
• PHP 7.4+
• An Anthropic API key

The key can be configured through the connector flow or supplied as ANTHROPIC_API_KEY in the server environment or WordPress configuration, depending on the implementation.

Appropriate uses

Claude is a practical choice for long-form editorial assistance, analysis, structured reasoning, document transformation and workflows requiring function calls. The precise model should be selected according to cost, latency, context and quality requirements rather than by provider name alone.

Google connector

The AI Provider for Google connects WordPress to Gemini and Google image-generation models. The current WordPress.org release supports Gemini text generation, Imagen image generation, function calling, automatic provider registration and dynamic model discovery.[^11]

Current capabilities

• Text generation with compatible Gemini models
• Multimodal text-and-image input where supported
• Image generation with Gemini and Imagen-compatible models
• Aspect-ratio controls for supported Gemini image generation
• Function calling
• Dynamic model discovery

Requirements

• WordPress 6.9 with the separate PHP AI Client package, or WordPress 7.0+
• PHP 7.4+
• A Gemini API key obtained through Google AI Studio

The connector documentation supports configuration through GOOGLE_API_KEY, while WordPress 7 also provides the centralized connector configuration experience.[^5][^11]

Appropriate uses

Gemini is relevant for multimodal content workflows, Google-centered technology stacks and image-generation use cases. Model capability must still be checked explicitly because not every Gemini model supports every combination of text, image, tools or structured output.

OpenAI connector

The AI Provider for OpenAI connects WordPress to OpenAI's text, image, embedding and speech models. Version 1.2.0 supports GPT text generation, DALL·E and GPT Image generation, function calling, web search, embeddings, image editing and text-to-speech.

Current capabilities

• Text generation with compatible GPT models
• Image generation with DALL·E and GPT Image models
• Image editing and refinement from reference images
• Function calling
• Web search on compatible models
• Embedding generation, including batch inputs and configurable dimensions
• Text-to-speech
• Discovery of available text, image and speech models
• Preservation of reasoning items and token usage across multi-turn Responses API calls

Version 1.2.0, released on September 21, 2026, added text-to-speech and improved multi-turn reasoning preservation. Version 1.1 had already added embeddings, fine-tuned model support and image editing.

Requirements

• WordPress 6.9 with the separate PHP AI Client package, or WordPress 7.0+
• PHP 7.4+
• An OpenAI API key

The key can be provided through the connector configuration or the OPENAI_API_KEY environment variable or constant.

Appropriate uses

OpenAI currently provides the broadest mix of modalities through this connector: text, web-assisted responses, embeddings, images, image editing and speech. That breadth can simplify a multi-modal WordPress implementation, although each operation still depends on the capabilities and commercial terms of the selected model.

Local models with Ollama

The WordPress AI Client architecture is not limited to the three featured cloud providers. The AI Provider for Ollama can connect WordPress to models running locally, on a private server or through a remote Ollama service.[^12]

Its advertised capabilities include text generation, embeddings, supported image generation, function calling, structured JSON output, typed decisions and automatic model discovery. A local installation can operate without a cloud API key, but it still requires adequate infrastructure and an Ollama instance reachable from the WordPress server.[^12]

Ollama can be useful when an organization wants stronger control over model hosting or wishes to avoid per-request cloud billing. It is not automatically private merely because the connector is called "local": privacy depends on where the model server runs, who can access it, whether prompts are logged and how the endpoint is secured.

Provider overview

No connector is universally best. A site can install more than one and select providers according to the operation, but that increases governance, testing and cost-control requirements.

The canonical AI plugin

The provider connectors are infrastructure. The canonical AI plugin is the component that gives editors and administrators ready-to-use features.

Version 1.4.0 was released on October 5, 2026. It requires WordPress 7.0.3+, PHP 7.4+ and the Block Editor. The Classic Editor and non-block editing experiences are not supported.[^7][^6]

Editorial and content features

• Alt Text Generation: generates image descriptions and integrates with the Image block's decorative-image setting.
• Content Classification: suggests tags and categories.
• Content Resizing: shortens, expands or reformulates selected block content.
• Content Summarization: creates summaries and supports bulk generation for posts and pages.
• Content Translation: translates paragraph and heading blocks and can optionally translate the title.
• Editorial Notes: reviews content block by block for accessibility, readability, grammar and SEO.
• Editorial Updates: applies accepted editorial suggestions.
• Excerpt Generation: proposes post excerpts.
• Image Generation and Editing: creates or edits images from the editor and Media Library.
• Meta Description Generation: proposes metadata and integrates with supported SEO plugins.
• Slug Generation: proposes editable SEO-friendly permalinks.
• Title Generation: produces headline ideas.
• Type Ahead: displays contextual ghost-text suggestions while writing.

Administration and moderation

• Comment Moderation: analyzes sentiment and toxicity and calculates a 0–1 value score reflecting how substantive and on-topic a comment appears.
• Suggest Reply: drafts responses for moderators.
• Abilities Explorer: displays registered abilities, schemas and test controls.
• AI Request Logging: records requests and provider errors for debugging and observability.
• Connector Approvals: requires administrator approval before plugins and themes use configured connectors.
• Custom Abilities: enables the plugin's general-purpose WordPress abilities through an explicit opt-in.
• Dashboard Widgets: summarizes provider status and available AI capabilities.
• Guidelines: lets compatible abilities respect reusable editorial guidance.
• Key Encryption: optionally encrypts provider API keys at rest.

All these features are opt-in experiments. WordPress recommends staging tests because the plugin is experimental and functionality may change.[^6]

Markdown feeds

AI 1.4 adds Markdown output at /feed/markdown/ and for individual posts and pages through ?output_format=markdown. The objective is to make WordPress content easier for AI systems and other Markdown consumers to parse.[^7][^6]

Markdown availability can reduce extraction friction, but it does not guarantee indexing, inclusion in model training, citation in an answer engine or improved rankings. Publishers should also consider whether making clean full-text Markdown versions available aligns with their content licensing and access strategy.

Embeddings

Version 1.4 adds a storage and CRUD layer for vectors in a new wpai_embeddings table, records the provider and model used for each vector, and introduces vector similarity and ranking classes. This creates infrastructure for semantic search, recommendations, retrieval-augmented generation and content matching, but it is not yet a turnkey enterprise search product.[^6][^7]

Function calling does not equal site control

The connector descriptions advertise function calling. This means a compatible model can return a structured request to invoke a function that the application has explicitly defined. It does not mean the model receives unrestricted access to WordPress.

For a safe function-calling workflow, a plugin must:

1. Define the available function or tool. 2. Describe its input schema. 3. Validate all model-supplied arguments. 4. Check the current user's permissions. 5. Execute the server-side callback. 6. Sanitize and log the result. 7. Return an appropriate response to the model. 8. Require human confirmation for consequential actions where appropriate.

For example, a plugin could expose create_article_draft, but it should not expose arbitrary PHP execution or direct database access. The permission boundary belongs to WordPress code, not to the model's instructions.

Abilities API

The Abilities API, available since WordPress 6.9, provides a registry for self-describing WordPress capabilities. Each ability has a namespace and name, human-readable description, JSON Schema inputs and outputs, an execution callback and a permission callback.[^8]

The API creates a reusable contract. The same ability can potentially be invoked by WordPress code, REST clients, an administrative interface or an MCP-compatible agent without redefining the business operation for each channel.

Registered abilities are not automatically exposed through REST. Exposure must be enabled, authenticated users are required, and each ability's permission callback still controls execution. Read-only, mutating and destructive operations use different HTTP methods in the REST interface.[^10]

WordPress 7.1 expanded the API with lifecycle filters, validation controls, public exposure metadata and ability discovery improvements, making it more practical for integrations and protocol adapters.[^13][^14]

MCP Adapter

The MCP Adapter bridges the Abilities API to Model Context Protocol. Version 0.7.0 is now available in the WordPress plugin directory, but it remains experimental while the team works toward a stable 1.0 release.[^3]

What it provides

• Conversion of WordPress abilities into MCP tools, resources and prompts
• A default MCP server
• Support for custom MCP servers
• HTTP and STDIO transports
• Permission checking
• Error handling and observability hooks
• Support for the 2025-11-25 and 2026-07-28 MCP schema revisions

The public announcement identifies the default endpoint as:

/wp-json/mcp/mcp-adapter-default-server

The adapter supplies infrastructure, not a complete chat product or agent UI. A compatible MCP client must connect to the site and authenticate as a WordPress user.[^3]

Exposure and permissions

Abilities must be public before the default MCP server exposes them. Even then, their existing permission callbacks and WordPress capability checks continue to apply. The adapter can therefore make an ability discoverable without bypassing its authorization rules.[^10][^3]

This is an essential distinction: MCP is a transport and discovery layer, not an authorization system that replaces WordPress permissions.

What the adapter does not include

• An AI model
• Provider credentials
• A built-in Claude or ChatGPT interface
• Automatic WooCommerce abilities
• A complete end-to-end site-management agent
• Automatic exposure of private abilities

The adapter can expose only what WordPress Core or installed plugins have registered and deliberately made public.[^3]

Can Claude or ChatGPT manage WordPress?

Yes, but not merely because an Anthropic or OpenAI connector is installed.

There are two different directions:

WordPress calling Claude or OpenAI

A WordPress plugin uses the AI Client and a provider connector to send prompts to a model. This is appropriate for generation, classification, translation, summarization and model-assisted internal workflows.

Claude or ChatGPT calling WordPress

An external MCP-compatible agent connects to the MCP Adapter, discovers public abilities and invokes only those for which the authenticated WordPress user has permission. This is appropriate for controlled agentic administration.[^3]

A production implementation could expose narrowly scoped abilities such as:

• Retrieve a post by ID
• Search published articles
• Create a draft, but not publish it
• Propose metadata changes
• Upload a validated image
• Add an internal editorial note
• Retrieve WooCommerce stock information
• Prepare a product update for human approval

The safer pattern separates proposal, drafting, approval and execution rather than granting a model unrestricted write access.

Can an agent edit WooCommerce products?

Not with the provider connectors or MCP Adapter alone.

WooCommerce operations must be exposed by WooCommerce itself, another plugin or custom code as carefully designed abilities or internal function-calling tools. A useful implementation might provide:

• commerce/get-product
• commerce/search-products
• commerce/create-product-draft
• commerce/propose-description-update
• commerce/update-stock
• commerce/update-price
• commerce/publish-product

These operations should not all share the same permission level. Reading product data is lower risk than changing inventory, pricing or publication state.

A production WooCommerce agent should include:

• Separate read and write abilities
• WordPress capability checks for every operation
• Strict JSON Schema validation
• Product and variation identifiers instead of fuzzy matching where possible
• Idempotency protection for repeated calls
• Audit logs containing the authenticated user, agent, ability, inputs and outcome
• Rate limits and spending controls
• Human approval for price, stock, tax, shipping and publication changes
• A rollback or recovery process

The recommended first implementation is a read-and-propose agent: it reads product information, drafts proposed changes and creates a reviewable task. Direct autonomous publishing should be introduced only after the organization has validated permissions, logs, error handling and rollback procedures.

Installation recipes

Editorial assistant inside WordPress

Use this stack when editors want AI functions in Gutenberg:

1. Update to WordPress 7.0.3 or later. 2. Install one or more provider connectors. 3. Configure credentials in Settings → Connectors. 4. Install and activate the canonical AI plugin. 5. Open Settings → AI. 6. Enable only the required experiments. 7. Approve connector access if Connector Approvals is enabled. 8. Test generation, errors, costs and editorial review on staging.

No MCP Adapter is required when all model calls originate inside WordPress.

External agent with read-only access

Use this stack when an MCP-compatible assistant needs to inspect a site:

1. Run WordPress 6.9+; WordPress 7.1+ is preferable for the expanded Abilities API. 2. Install the MCP Adapter. 3. Enable or create only the necessary read abilities. 4. Mark only the intended abilities as public. 5. Create a dedicated WordPress user with minimum privileges. 6. Configure the MCP client with the server endpoint and supported authentication. 7. Confirm that unauthorized abilities are neither executable nor unnecessarily exposed. 8. Review logs and revoke credentials after testing.

A provider connector is not required merely to expose WordPress through MCP; the model runs in the external client. It is required only if WordPress itself also needs to call a provider.

Controlled publishing agent

A write-capable agent requires custom abilities or a plugin supplying them. Begin with draft creation, require human review, and separate publishing into a more privileged ability. Avoid using a full administrator account for routine agent actions.

Security and governance

AI integration introduces two distinct trust boundaries: content sent from WordPress to a model provider, and actions sent from an external agent to WordPress.

Data sent to providers

Before enabling a feature, identify what its prompt may contain. Depending on the workflow, requests can include post content, unpublished drafts, comments, usernames, media, metadata or internal instructions. The applicable provider contract, retention policy, processing location and training policy should be checked rather than inferred from the connector.

For GDPR-sensitive sites, document:

• The provider and model used
• The categories of personal or confidential data transmitted
• The legal basis and processor relationship
• Retention and deletion terms
• International data transfers
• Whether provider-side training is disabled
• Who can trigger requests
• Where prompts, responses and request logs are stored

Do not send passwords, private keys, authentication tokens, payment details, health information or unnecessary personal data to a model.

Prompt injection

Any workflow that processes external content—web pages, comments, uploaded documents, product feeds or emails—must treat that content as untrusted. Text can contain instructions intended to manipulate the model into disclosing information or invoking an inappropriate tool.

Mitigations include:

• Give the agent the smallest possible tool set.
• Keep read and write operations separate.
• Validate arguments independently of the model.
• Never let a prompt override WordPress permissions.
• Require confirmation for consequential actions.
• Prevent tools from returning secrets the agent does not need.
• Log tool calls and investigate unusual sequences.

Credentials

Use separate API keys for development, staging and production. Apply provider spending limits where available, rotate keys, restrict access to configuration files and ensure secrets are not committed to version control.

For MCP access, use a dedicated WordPress account and revokeable credentials. The account's role should match the abilities it must execute, not the broadest role convenient during development.

Production readiness

The canonical AI plugin and MCP Adapter are explicitly experimental. WordPress recommends staging tests for the AI plugin, while the MCP Adapter team is still gathering feedback before version 1.0.[^6][^3]

For production use:

• Pin tested versions.
• Maintain backups and rollback procedures.
• Test updates before deployment.
• Monitor provider errors, latency and costs.
• Review generated content for factual, legal and brand risks.
• Disable unused experiments, connectors and abilities.
• Verify that logs do not retain sensitive prompts longer than necessary.

Costs

The WordPress plugins are open-source and free to install, but cloud-provider usage is generally billed separately. Cost depends on the model, input and output volume, images, embeddings, speech, web-search tools and retry behavior.[^6]

A governance plan should define:

• Approved providers and models
• Default model by task
• Maximum input and output sizes
• Monthly budgets and alerts
• Rules for expensive image or reasoning models
• Cache and reuse policies
• Logging and cost attribution by site or feature

Local Ollama models avoid a per-request cloud invoice but still consume hardware, electricity, administration time and hosting capacity.[^12]

Choosing the right architecture

Current limitations

The current official stack is substantial, but it is not a fully autonomous site-management product.

• The AI plugin supports Gutenberg, not the Classic Editor.[^6]
• Provider connectors supply model access, not finished business workflows.
• Model support varies by provider and can change as APIs evolve.
• Content Translation currently focuses on paragraph and heading blocks; documented testing has identified gaps for lists, tables and shared terminology glossaries.[^15]
• Markdown feeds improve machine readability but do not guarantee AI citations or search visibility.
• Embedding infrastructure is available, but complete semantic-search and RAG experiences still require implementation.
• The MCP Adapter has no complete agent UI and remains pre-1.0 experimental software.[^3]
• Public ability exposure remains opt-in, and permissions must be designed correctly.
• WooCommerce management requires additional abilities or custom development.
• WordPress 7.2 roadmap items—including seamless MCP access, WebMCP experimentation, agent identity, more abilities, semantic search and streaming—are being pursued without a guarantee that they will enter Core in 7.2.[^16]

What WordPress 7.2 may change

The WordPress 7.2 roadmap continues AI work primarily through canonical plugins while requiring features to demonstrate adoption and practical value before consideration for Core.[^16]

Current areas of development include:

• Additional read and write abilities
• Easier MCP setup and ability-exposure controls
• WebMCP experiments linked to the active browser interface
• Auditable agent identity and delegated access
• Reusable editorial knowledge and guidelines
• Embeddings and semantic search
• Streaming AI responses

These items should be described as roadmap work, not as finished WordPress 7.2 features, until they appear in a release candidate or final release.

Recommended agency architecture

For an agency managing several WordPress sites, a controlled reference architecture is preferable to installing every experiment everywhere.

Baseline

• WordPress 7.0.3+ with current security updates
• One approved connector per primary workload, with a second provider only where justified
• Environment-based secret management
• Canonical AI plugin enabled only on sites that need editor-facing features
• Connector Approvals and request logging
• Staging validation before every AI or MCP update

Agentic tier

• WordPress 7.1+
• MCP Adapter on sites requiring external agents
• Dedicated agent users and credentials
• Separate read, propose, draft and publish abilities
• Human approval for public or commercial changes
• Central logs, alerts and incident revocation process

Commerce tier

• Purpose-built WooCommerce abilities
• No direct generic database or PHP tools
• Price and inventory limits
• Explicit store, currency and product identifiers
• Approval for high-impact updates
• Post-action verification against WooCommerce data

This architecture keeps the provider layer replaceable while treating WordPress permissions and business rules as the final authority.

Frequently asked questions

Are Claude, Gemini and OpenAI built directly into WordPress?

No. WordPress Core includes the provider-independent AI Client and connector framework, but provider implementations remain separate plugins. A provider connector and credentials must be configured before WordPress can call that service.[^5][^1]

Does installing the OpenAI connector add ChatGPT to WordPress?

No. It creates a standardized connection to compatible OpenAI models. Install the canonical AI plugin for its ready-made editorial features, or install/build another plugin that uses the AI Client.

Do these plugins require a ChatGPT, Claude or Gemini consumer subscription?

They require appropriate API access, which is generally distinct from a consumer chatbot subscription. Obtain and bill API credentials through the provider's developer platform.

Can a site use several providers?

Yes. Multiple connectors can coexist, and an application can request a capability or select an appropriate provider and model. Organizations should nevertheless control provider access, costs and data flows.

Can models be changed without rewriting a plugin?

That is a central benefit of the AI Client, provided the plugin requests standardized capabilities rather than depending on provider-specific behavior. Differences between models can still require testing and prompt adjustments.[^1]

Is MCP required for AI generation in Gutenberg?

No. Gutenberg features use the AI Client and a provider connector. MCP is relevant when an external compatible agent needs to discover and invoke approved WordPress abilities.

Can MCP bypass WordPress permissions?

It should not. Public exposure is separate from authorization, and permission callbacks run for the authenticated user. A badly designed custom ability can still create risk, so permissions and input validation must be reviewed carefully.[^10][^3]

Can an agent publish posts automatically?

Only if a write ability has been installed or developed, exposed through MCP and authorized for the authenticated user. A safer design allows draft creation but reserves publishing for a human or more privileged approval step.

Does the official stack support WooCommerce automatically?

No. WooCommerce operations require abilities or function-calling tools supplied by WooCommerce, another plugin or custom development.

Is the canonical AI plugin production-ready?

WordPress describes it as experimental and recommends testing on staging. Use in production should therefore be selective, monitored and reversible.[^6]

Does WordPress send site data to AI providers by default?

No. Core provides infrastructure, but a connector must be configured and code or an enabled feature must explicitly initiate a request.[^9]

The practical takeaway

WordPress now has a coherent official AI stack rather than three isolated provider plugins. The AI Client standardizes requests; connectors supply access to Claude, Gemini, OpenAI and other systems; the canonical AI plugin delivers editor-facing experiments; the Abilities API describes WordPress operations; and the MCP Adapter exposes approved operations to external agents.[^8][^1][^3]

For most publishers, the starting point is WordPress 7.0.3+, one connector and a small set of AI plugin features. For agencies and commerce sites, the strategic opportunity lies in abilities and MCP—but only with narrow permissions, dedicated identities, complete logging and human approval for consequential actions.

References

1. Introducing the AI Client in WordPress 7.0 2. WordPress 7.0 Field Guide 3. WordPress MCP Adapter 0.7.0 launches on the WordPress Plugin Directory 4. dev-notes-7-0 – Make WordPress Core 5. Introducing the Connectors API in WordPress 7.0 6. AI – WordPress plugin 7. What's new in AI 1.4.0 (5 October 2026)? 8. Abilities API – Common APIs Handbook 9. Proposal for merging WP AI Client into WordPress 7.0 10. REST API endpoints – Common APIs Handbook 11. AI Provider for Google – WordPress plugin 12. AI Provider for Ollama – WordPress plugin 13. What's new for developers? (August 2026) 14. 7.1.0 | Page 5 of 8 | Developer.WordPress.org 15. AI Contributor Weekly Summary – 26 August 2026 16. Roadmap to 7.2 – Make WordPress Core