Laravel agents now speak MCP natively — and how they shipped it is the real story
The Laravel AI SDK can now point an agent at any MCP server. The interesting part isn’t that it works — it’s that the team refused the shortcut.
The most-requested feature for the Laravel AI SDK landed. Point an agent at a server — Nightwatch, GitHub, Notion, or one you run locally — drop its tools into the agent’s existing <code>tools()</code> method, and the model uses them as if hand-written. STDIO for local servers, streamable HTTP for remote. Bearer tokens and full OAuth (PKCE, redirects, refresh handled for you). The tool list is cacheable, and you can test the whole path with the SDK’s fake layer.
If you build agents in PHP, the glue you’ve been writing by hand is now framework-grade. MCP tools and your own tools sit in one array and the model can’t tell them apart — that deletes a chunk of boilerplate and makes Laravel a credible home for production agents, not just demos.
I’ve built MCP servers in PHP by hand (<code>helpmatch-mcp</code>, my own <code>eduardocruz</code> server), so the call I’d defend is the one people skim past: they did not bolt a quick wrapper inside <code>laravel/ai</code>. They put the client in <code>laravel/mcp</code>, next to the server code — so a queued job or console command can hit an MCP server with no agent in sight. That’s the line between a demo feature and infrastructure. This is the "Laravel 2 in 2012" moment for AI.
EC TV is written by Eduardo Cruz — a senior Laravel engineer who ships production AI agents and MCP servers.
Work with me → Read the deep-dives