Calling a language model from Laravel is a POST request, and for a one-off you can write it with the Http client. Once you need structured output, tool calling, streaming to the browser, retries, and the ability to swap providers, an SDK earns its place. This guide covers the Laravel AI ecosystem in 2026 and how to use each capability, with Prism as the concrete example.

The 2026 landscape

• ✓Prism (prism-php/prism): provider-agnostic, one API across Anthropic, OpenAI, Mistral, Ollama and others. Text, structured output, tools, streaming, embeddings.
• ✓Official provider SDKs: the vendor packages when you want provider-specific features and are not planning to switch.
• ✓Laravel MCP: exposes your application as tools an MCP client (including an LLM agent) can call, the other direction from calling a model.

Text generation

The basic call: pick a provider and model, set a system prompt, send the user prompt. Keep the model choice in config so you can change it without touching code.

use Prism\Prism\Prism;

$response = Prism::text()
    ->using('anthropic', config('services.llm.model')) // e.g. claude-sonnet-5
    ->withSystemPrompt('You are a concise support assistant.')
    ->withPrompt($question)
    ->asText();

return $response->text;

Structured output

When you need data, not prose, describe a schema and get back a typed array that has already been validated against it. This is the right tool for extraction (pull fields out of an email), classification (route a ticket), and any case where the next step is code, not a human reading text.

use Prism\Prism\Schema\{ObjectSchema, StringSchema, EnumSchema};

$schema = new ObjectSchema('ticket', 'A triaged support ticket', [
    new StringSchema('summary', 'One sentence summary'),
    new EnumSchema('category', 'The team it belongs to', ['billing', 'technical', 'account']),
    new EnumSchema('priority', 'Urgency', ['low', 'medium', 'high']),
], ['summary', 'category', 'priority']);

$response = Prism::structured()
    ->using('anthropic', config('services.llm.model'))
    ->withSchema($schema)
    ->withPrompt($emailBody)
    ->asStructured();

$ticket = $response->structured; // ['summary' => ..., 'category' => ..., 'priority' => ...]

Tool calling, kept safe

A tool lets the model call a function in your code: look up an order, check stock. The SDK handles the round trip, but you stay on the proposing side only. The model returns a tool name and arguments; your closure validates them against the current user's permissions and your business rules, then runs or refuses. The model never touches the database directly.

use Prism\Prism\Tool;

$orderStatus = Tool::as('order_status')
    ->for('Get the status of an order for the current user')
    ->withStringParameter('order_id', 'The order reference')
    ->using(function (string $orderId) use ($user) {
        $order = $user->orders()->find($orderId); // scoped to the user
        return $order ? "Status: {$order->status}" : 'Order not found';
    });

$response = Prism::text()
    ->using('anthropic', config('services.llm.model'))
    ->withTools([$orderStatus])
    ->withMaxSteps(3)
    ->withPrompt($question)
    ->asText();

Streaming to the browser

A generation call takes seconds; stream the tokens so the interface responds immediately. The SDK exposes the response as an iterable; wrap it in a Laravel streamed response over Server-Sent Events and consume it with EventSource on the front end.

return response()->stream(function () use ($question) {
    $stream = Prism::text()
        ->using('anthropic', config('services.llm.model'))
        ->withPrompt($question)
        ->asStream();

    foreach ($stream as $chunk) {
        echo 'data: ' . json_encode(['text' => $chunk->text]) . "\n\n";
        ob_flush();
        flush();
    }
}, 200, ['Content-Type' => 'text/event-stream', 'X-Accel-Buffering' => 'no']);

Embeddings and pgvector

The same SDK generates embeddings, which you store in a pgvector column next to your relational data for a lightweight RAG setup, no separate vector database. Embed your documents at ingestion, embed the question at query time, and run a similarity search in SQL.

SDK abstraction or provider SDK directly?

Cost, caching, rate limiting, fallback

• ✓Cache answers to frequent or identical prompts so the same question is not paid for twice.
• ✓Rate-limit per user and per API key, and return a clear 429.
• ✓Track tokens and cost per feature, attributed to a user or tenant.
• ✓Have a non-AI fallback (a help page, a form) for when the provider times out or errors.

An AI SDK does not add intelligence; it removes plumbing. It gives you one shape for text, structured output, tools, streaming and embeddings across providers, so the interesting work, the prompt, the context, the tool boundaries, is where your effort goes.