The Laravel team released Laravel AI SDK 1.0 on September 23, 2026. The first-party package provides one API for working with AI providers in a Laravel application. Laravel first announced the SDK in February.

terminal
composer require laravel/ai

Classification is now a separate capability alongside text, images, audio, and embeddings. It handles decisions such as support-ticket routing and comment flagging through Jev models from TypeSafe. The announcement describes Jev as returning answers in milliseconds at a fraction of the cost of traditional LLMs. Boolean and Choice questions return typed results, while Score produces a value from 0.0 to 1.0. The `Str::decide` macro handles a single yes-or-no question and accepts a confidence threshold. Classification currently supports TypeSafe and OpenRouter, with more providers planned.

The SDK can consume requests and stream responses through the Vercel Chat and AG-UI protocols. A Vercel integration can create a chat with `Vercel::chat($request)`, pass its history to an agent, and stream the response with the selected protocol. The route also updates conversation history in the database and handles submitted tool approvals. AG-UI clients such as CopilotKit use `usingAgentUserInteractionProtocol()`. `Vercel::toUiMessages()` converts stored messages back to the client format after a page reload.

Tools that implement `Approvable` and use the `InteractsWithApprovals` trait can pause an agent until a person responds. Each pending call includes the arguments selected by the model. The user can approve the call, reject it with a reason visible to the model, or edit its arguments before execution. The workflow works with prompt, stream, queue, and broadcast methods. Laravel News previously covered the feature when it arrived in July.

Agent middleware now runs at every generation step. Each step is represented by `PendingStep`, which exposes methods such as `withModel`, `withTools`, `withoutTools`, and `withMaxTokens`. This allows an application to remove an expensive tool after the agent has used it once. `ToolSearch` loads infrequently used tools only when a prompt requires them, with support for OpenAI and Anthropic. The `CodeExecution` provider tool runs code in provider sandboxes on Anthropic, OpenAI, Azure, Gemini, and xAI.

Conversation messages now use one `steps` JSON column with one entry per round trip. Each tool result is stored with the call that produced it. Usage fields named `promptTokens` and `completionTokens` become `inputTokens` and `outputTokens`, and the values include the provider's complete counts across providers.

Version 1.0 contains breaking changes involving conversation storage, agent middleware, token usage, and stream protocols. Applications that query the `tool_calls` or `tool_results` columns with raw SQL must move those queries to `steps`. The upgrade guide includes a backfill migration that must run once before deployment.

Laravel recommends using Laravel Boost to assist with the upgrade. The process installs `laravel/boost` as a development dependency, runs `php artisan boost:install`, and then uses `/upgrade-ai-sdk-v1` in Claude Code, Cursor, OpenCode, Gemini, or VS Code. Boost guides the assistant through the upgrade changes one at a time.