Laravel MCP is the 2026 Laravel/AI topic that is easy to over-buy. MCP (Model Context Protocol) lets an external agent — ChatGPT, Claude, Cursor — call tools you define instead of guessing from a PDF of your schema.
That is powerful inside a business system. It is also a new door into your data. This article is when MCP belongs in a Philippine Laravel app, and when it is theater.
Laravel’s own split is useful: the AI SDK builds features inside the app; Laravel MCP exposes the app outward. We covered the inward path in Laravel 13 AI SDK for business apps.
What MCP is in plain language
Without MCP, every AI product invents its own integration. With MCP, you run a server in Laravel that advertises tools and resources: “list today’s unpaid invoices,” “find a court slot,” “draft a reply using this policy.”
A compatible client can then call those tools with your auth story, not with a pasted API key in a random custom GPT.
That is the appeal for ops: the agent uses the same rules as your admin, if you wired it that way. If you did not, you built a second, sloppier API.
When MCP is worth the work
Consider a server when:
- Staff already live in an AI client and in your Laravel admin, and the copy-paste is the job
- You can name five tools, not fifty — each maps to an existing policy (who may refund, who may see PII)
- You have auth: Sanctum/Passport, roles, audit logs, rate limits
- The alternative is giving an intern production DB credentials “so Claude can help”
MCP is a poor first project when you do not yet have a system of record. An agent cannot book honestly if bookings live in three sheets. Fix the core — business systems vs websites — then expose tools.
A shape that does not scare us
| Layer | Job |
|---|---|
| Domain + API | Real states, permissions, webhooks |
| AI SDK (optional) | In-app drafts your UI already shows |
| MCP server | Narrow tools for trusted clients |
| Auth + audit | Every tool call is a user (or service) you can revoke |
Do not put the MCP endpoint on the public internet with a long-lived token in a chat. Treat it like admin API access. The composable-integration habits in connecting tools without replatforming apply: least privilege, rotation, logging.
Tools we would expose first (and last)
First (read-mostly, high value):
- Search orders/bookings the caller is allowed to see
- Summarize this record for a standup
- Fetch the written policy snippet the agent must follow
Later (writes, after months of clean reads):
- Create a draft invoice or draft booking in
pending_review - Never: silent refunds, silent price edits, silent user deletes
If a vendor demo writes to production from ChatGPT on day one, that is a sales demo, not a design.
MCP is not “our app is agentic”
Agentic means multi-step action. MCP is a pipe. You can have MCP and still require a human click in Laravel. You should.
Developer-facing MCP (helping Cursor see your schema) is a productivity choice. Product-facing MCP (helping a GM ask ChatGPT for yesterday’s no-shows) is a security and product choice. Do not fund them as one line item.
Upgrade note
You do not need Laravel 13 to think about this, but 13 is where the first-party AI story is concentrated. If you are still on 12, the framework conversation is upgrade now or wait — separately from “should agents touch our books.”
Straight next step
If you want agents to talk to a real Philippine business system — not a toy todo list — describe the three questions you wish ChatGPT could answer from your admin. We will say which should be an MCP tool, which should be a report in Laravel, and which should stay a human.
