silverstripe-ai

Core AI platform integration for SilverStripe, built on Symfony AI. Provides a unified AIClient service and a /ai/generate endpoint that works with OpenAI, Anthropic, Google (Gemini & Vertex AI), Azure OpenAI, Mistral, Ollama, OpenRouter, and any OpenAI-compatible endpoint.

Requirements

• SilverStripe Framework ^6
• PHP ^8.1

Installation

composer require xddesigners/silverstripe-ai

Configuration

Add the following to your .env file:

AI_PLATFORM_TYPE="openai"   # see the table below
AI_MODEL="gpt-4o-mini"
AI_API_KEY="sk-xxx"

Supported platforms and models

Platform type	Example models	Extra environment variables
openai	gpt-4o, gpt-4o-mini, gpt-3.5-turbo	—
anthropic (claude)	claude-sonnet-4-6, claude-haiku-4-5	—
gemini (google)	gemini-2.0-flash, gemini-1.5-pro	—
azure	your deployment name	AI_PLATFORM_BASE_URL (resource endpoint), AI_AZURE_DEPLOYMENT, AI_AZURE_API_VERSION (default 2024-10-21)
vertex	gemini-1.5-pro	AI_VERTEX_LOCATION, AI_VERTEX_PROJECT (key optional; ADC supported)
mistral	mistral-large-latest, mistral-small-latest	— (native bridge; AI_PLATFORM_BASE_URL only used for the generic fallback)
ollama	any locally pulled model	AI_PLATFORM_BASE_URL (optional; defaults to http://localhost:11434)
openrouter	any OpenRouter slug, e.g. anthropic/claude-sonnet-5	AI_PLATFORM_BASE_URL (optional; set to route in-region — see below)
generic	any model on the endpoint	AI_PLATFORM_BASE_URL (required) — any OpenAI-compatible endpoint

mistral uses the native Symfony Mistral bridge (bundled), falling back to Mistral's OpenAI-compatible EU endpoint if that package is removed. ollama and generic reach their providers through the OpenAI-compatible generic bridge, so no extra package is needed.

EU data residency

Keep inference inside the EU by pointing an OpenAI-compatible platform at an in-region endpoint via AI_PLATFORM_BASE_URL:

# OpenRouter EU entry point (Business plan). Same key and model slugs; requests are processed in the EU.
AI_PLATFORM_TYPE="openrouter"
AI_PLATFORM_BASE_URL="https://eu.openrouter.ai/api"
AI_MODEL="anthropic/claude-sonnet-5"
AI_API_KEY="sk-or-xxx"

Other EU-friendly options: mistral (Mistral’s EU platform, the default endpoint), azure in an EU region, or ollama/generic against a model you host in the EU — data then never leaves your infrastructure.

Optional YAML configuration

You can override AIClient defaults in YAML:

XD\SilverstripeAI\Services\AIClient:
  max_text_length: 5000
  max_instructions_length: 1000
  default_instructions: 'You are a helpful assistant and SEO expert.'
  # Cost estimates are best-effort. Add or override per-model rates (USD per 1,000,000 tokens);
  # keys may be a bare model name or an OpenRouter-style "vendor/model" slug.
  model_pricing:
    'anthropic/claude-sonnet-5':
      input: 3.00
      output: 15.00

Usage

In PHP

Inject or instantiate AIClient and call one of its two methods:

use XD\SilverstripeAI\Services\AIClient;

$client = AIClient::create();

// Generate a plain text response
$result = $client->generateText('Write a short intro for our homepage.');

// Generate a keyed set of fields from context data
$result = $client->generateFields(
    context: ['Title' => 'My Page', 'Content' => 'Existing body text…'],
    fields: ['Title', 'Content', 'MetaDescription'],
    instructions: 'Improve the content for SEO.'
);
// Returns: ['Title' => '…', 'Content' => '…', 'MetaDescription' => '…']

Via the HTTP endpoint

The module registers a /ai/generate route that accepts POST requests.

Text mode

POST /ai/generate
mode=text&text=Write+a+short+intro&instructions=Keep+it+friendly

Response:

{ "result": "…" }

Fields mode

POST /ai/generate
mode=fields
&context[Title]=My+Page
&context[Content]=Existing+body+text
&fields[]=Title
&fields[]=Content
&fields[]=MetaDescription
&instructions=Improve+for+SEO

Response:

{ "fields": { "Title": "…", "Content": "…", "MetaDescription": "…" } }

Error responses

Status	Meaning
400	Missing or invalid input
429	Rate limit exceeded (includes retry_after_seconds)
500	AI generation failed

Usage logging & the AI Usage admin

Every request is logged to AIRequestLog (platform, model, mode, token counts and estimated cost) and can be browsed under the AI Usage CMS section.

Token usage is captured for every supported provider. Native bridges (OpenAI, Anthropic, Gemini, Vertex AI, Mistral) report it through Symfony AI's token_usage metadata; for OpenAI-compatible endpoints that don't surface that metadata (OpenRouter, Ollama, the generic bridge) the counts are read from the response's usage block as a fallback. Cost is only estimated when the model is listed in model_pricing (see below) — otherwise the request is logged with token counts but no cost.

Who can see it

• Administrators always have access.
• Grant other groups access by ticking “Access to 'AI Usage' section” under Security → Groups → Permissions. Unlike most CMS sections, holding “Access to all CMS sections” does not reveal it.

Hiding it from clients

Set this in .env to hide the section entirely (menu + access), for everyone including admins:

AI_USAGE_ADMIN_DISABLED=1

Privacy & data processing

This module sends the text you pass to AIClient — page content, field values, prompts — to the third-party AI provider you configure. Treat that as a data-processing step:

• Personal data / GDPR: if that content can contain personal data, the AI provider acts as a processor. Choose the provider and region accordingly, record it in your processing register, and cover it in your privacy policy / DPA. For EU data residency, route through an in-region endpoint (see EU data residency above) — e.g. OpenRouter's eu.openrouter.ai, Mistral's EU platform, Azure in an EU region, or a self-hosted model.
• What is logged: AIRequestLog stores only metadata (platform, model, token counts, estimated cost and the requesting CMS member) — never the prompt or the response. Enabling the AI Usage admin does not create a store of processed content.
• Endpoint trust: your AI_API_KEY is sent as a bearer token to AI_PLATFORM_BASE_URL (for the openrouter / generic / mistral / ollama types). Point it only at endpoints you trust.
• Access & hardening: /ai/generate requires a logged-in CMS user (CMS_ACCESS), is POST-only, CSRF-protected and per-member rate limited. Do not run the site in dev mode in production — dev mode returns raw error detail to the client.

Extending the controller

By default the endpoint is POST-only and AIController::assertAccess() requires a logged-in member (returning Security::permissionFailure() otherwise). Subclass the controller to add your own permission logic — return an HTTPResponse to deny, or null to allow:

use SilverStripe\Control\HTTPResponse;
use SilverStripe\Security\Permission;
use SilverStripe\Security\Security;
use XD\SilverstripeAI\Controllers\AIController;

class MyAIController extends AIController
{
    protected function assertAccess(): ?HTTPResponse
    {
        if (!Permission::check('MY_AI_PERMISSION')) {
            return Security::permissionFailure($this);
        }
        return null;
    }
}

Then update the route in YAML:

SilverStripe\Control\Director:
  rules:
    'ai//$Action': 'MyAIController'

Suggested modules

• xddesigners/silverstripe-ai-assistant — adds a ready-to-use AI Assistant tab to CMS edit forms, letting editors generate and preview AI-written content for any configured fields.

License

BSD-3-Clause © XD Designers