I've been wary of AI coding assistants for a good while now. Not because I think they're a fad, they're clearly not, but because every one I tried had the same habit: confidently suggesting Laravel code that doesn't work. A Model::with() call using a relationship that doesn't exist. A migration for a column that's already there under a different name. Config keys from Laravel 9 in a Laravel 12 project. You spend more time correcting it than you'd have spent just writing the thing yourself.
So when Laravel Boost landed from the Laravel team, I was curious but not exactly hopeful. I've been running it on a real client project for the past few weeks, a booking system for a leisure company, built on Laravel 12 with Livewire 3 and Flux UI, and it's genuinely changed how I use Claude and Cursor day to day.
What it actually does
Boost is an MCP server. If you haven't come across MCP (Model Context Protocol) yet, think of it as a way for AI tools to ask your application questions instead of guessing. Install it with composer require laravel/boost --dev, run php artisan boost:install, and it wires itself up to whichever AI tool you're using, Claude Code, Cursor, VS Code Copilot, that sort of thing.
Once it's connected, the assistant can pull your actual route list, inspect your real database schema, run things through Tinker, and search documentation that matches the exact package versions you've got installed, not whatever version it was trained on. That last part matters more than it sounds. Half the wrong answers I used to get were because the model was quietly assuming Laravel 10 behaviour in a Laravel 12 app.
Where I actually noticed the difference
On the booking project, I asked Claude to add a cancellation window, stop customers cancelling within 24 hours of a slot. Before Boost, it wrote a migration adding a cancellation_cutoff column, completely missing that we already had a starts_at timestamp it could calculate against, and a cancellable_until field on the bookings table from an earlier feature. With Boost connected, it checked the schema first, found the existing columns, and used them. No new migration needed. That's a small thing, but it's the difference between a five-minute fix and a half hour of me untangling duplicate logic three weeks later.
It's also useful for the boring stuff. Asking it to check whether a route already exists before adding one, or to run a quick Tinker query to see what a relationship actually returns, rather than assuming. It feels less like a very confident intern and more like something that checks its work.
The bit I'm still cautious about
I don't let it run anything against a database with real client data without watching closely. Boost gives the AI read access to a lot of your app, and for freelance work under NDA, that's not a decision to make lightly. I had a conversation with the client before switching it on for that repo, mostly to be upfront about what tooling I use and why. They were fine with it, but I wouldn't assume every client would be, and I think that's worth checking rather than assuming.
It's also not magic. It still gets things wrong sometimes, particularly with anything Livewire-specific where the pattern isn't obvious from the schema alone. I still read every diff before I accept it. What's changed isn't that I trust the output blindly, it's that the starting point is closer to correct, so the review takes less time.
If you're already using an AI assistant day to day and haven't tried Boost, it's a quick install and free to try on a side project first. Point it at something small, ask it a question you already know the answer to, and see how close it gets. That's the real test.