Ask an AI assistant "how many of order 4471's items have shipped?" and it cannot answer, because the answer lives in your ERP and the assistant cannot see it. The Model Context Protocol, usually shortened to MCP, is the open standard that fixes this. Introduced in late 2024 and adopted across the major AI platforms through 2025 and 2026, it defines how an AI model discovers and calls tools exposed by a business system, with the business controlling exactly what is exposed.

If "AI agents" were the headline of 2025, MCP is the plumbing that makes them usable in 2026. This post explains it without the jargon and tells you what to ask your software vendors.

The problem MCP solves

Before MCP, connecting an AI assistant to a system meant a custom integration for every pair: this model with that CRM, that model with this ERP. Each one was a one-off, poorly documented and usually built with more access than it needed. MCP replaces this with a standard: a system runs an MCP server that publishes a list of tools (such as "look up order", "list overdue invoices", "create a support ticket") with typed inputs and descriptions. Any MCP-capable assistant can then use them.

What that enables in practice

  • Questions over live data: "Which customers in Ernakulam have invoices over 60 days?" answered from the real ledger, not a stale export.
  • Actions with approval: "Draft a reminder to each of them" where the assistant prepares the messages and a human approves sending.
  • Cross-system work: an assistant that reads a stock shortfall from the ERP, checks the supplier portal and drafts a purchase order, because each system speaks the same protocol.
  • Staff self-service: a sales rep asks the company assistant in WhatsApp or Slack for a customer's history instead of logging into three systems.

Why vendors are adding it

Enterprise ERP and CRM vendors shipped MCP support through 2025 and 2026 because customers asked for it: an AI assistant that cannot reach the system of record is a toy. For the custom and mid-market systems most Kerala businesses run, the question is whether your developer can add an MCP server to the application you already own. For a Laravel system, the answer is yes, and it is a modest project.

How an MCP server for a Laravel ERP works

  1. Choose the tools: start with five to ten read-only operations the business asks about every day. Reads before writes.
  2. Expose them: each tool is a thin layer over the service classes the application already has, with a JSON schema describing its inputs. Official and community MCP packages for PHP and Laravel handle the protocol.
  3. Authenticate every call: the assistant acts as a specific user with that user's permissions, through OAuth or a scoped token. The sales rep's assistant sees the sales rep's customers, nothing else.
  4. Log everything: every tool call, who triggered it and what it returned, in an audit table. This is also what DPDP compliance expects.
  5. Add writes with confirmation: creating or changing records goes behind an explicit human approval step in the first version.

The rules that keep it safe

  • Least privilege: tools expose the minimum fields. "Look up customer" returns name and order status, not the full record with ID proof.
  • No raw SQL tools: a tool that runs arbitrary queries is a breach waiting to happen, however convenient it looks in a demo.
  • Prompt injection awareness: data the assistant reads, such as a customer's email, can contain text trying to steer it. Keep write actions behind approval and never let tool output grant new permissions.
  • Rate limits and cost caps: an assistant in a loop can call a tool thousands of times. Limit it.
  • Private by default: the MCP server sits inside your network or behind your authentication, never on a public endpoint without auth.

Questions to ask your software vendor

  1. Does the system expose an MCP server, or a documented API we can wrap one around?
  2. Is access per user, with the same role permissions as the normal login?
  3. Is every AI call logged?
  4. Can we choose which AI provider connects, or is it locked to one?

Our ERP solutions and business automation work now includes MCP servers as a standard option, built on the Laravel AI tooling we described in Laravel 13 and PHP 8.5.

Frequently asked questions

Is MCP tied to one AI company?

No. It is an open specification with implementations across the major assistants and model providers, which is why building a server once is worthwhile.

Does the AI model get a copy of my database?

No. The model sees only what each tool call returns, for the user who made the request, and nothing is retained unless you choose a provider that retains it. Pick providers with clear data-handling terms.

How long does it take to add to an existing Laravel system?

A read-only server with a handful of tools, proper per-user authentication and audit logging is typically two to three weeks. Write actions with approval flows add one to two more.