For the last two years, the hard part of building anything useful with AI was not the model. It was the plumbing. Every time you wanted an assistant to read a Google Sheet, look up an order in your database, or post to WhatsApp, someone had to write custom glue code for that one connection. Ten tools meant ten integrations, each maintained by hand.
Model Context Protocol (MCP) is the standard that fixes this. It defines one common way for an AI model to discover and call external tools and data sources — so a connector you build once works with any model that speaks the protocol. The shift has been fast: by early 2026 the ecosystem reported roughly 97 million monthly SDK downloads and well over 5,000 public servers, and a Stacklok industry survey found about 41% of software organizations already running MCP servers in limited or broad production. It is moving from demo to default.
What MCP actually is
The simplest analogy is a port. Before USB, every device needed its own cable and its own driver. USB-C replaced that mess with one connector everyone agreed on. MCP does the same for AI: instead of bespoke code for each model-to-tool link, you expose your data and actions through an MCP server, and any MCP-capable client — a chatbot, an agent, an IDE — can use it.
A server publishes three kinds of things. Tools are actions the AI can take, like “create invoice” or “check stock”. Resources are data it can read, like a document or a database row. Prompts are reusable templates. The client asks the server what it offers, then calls those capabilities in a structured, predictable way.
Without MCP With MCP
---------------------- ----------------------------
Agent --custom code--> CRM Agent --+
Agent --custom code--> DB | one protocol
Agent --custom code--> Chat MCP client
Agent --custom code--> Sheets |
+-------+-------+--------+
N tools = N integrations CRM DB Chat Sheets
each built and maintained (each an MCP server, reusable)The payoff is leverage. Build an MCP server for your inventory system once, and it works with this year’s model and next year’s — and with your customer chatbot, your internal ops agent, and a developer’s coding assistant, all without rewriting the connection.
Why this matters for Indonesian businesses
Most companies here do not run one system. You might have a POS, a marketplace account, a WhatsApp line, an accounting app, and a spreadsheet that quietly holds the truth. The value of an AI agent is not in any single one of those — it is in connecting them. “A customer on WhatsApp asks about an order; check the POS, confirm stock, update the sheet, reply.” That is four systems in one sentence.
MCP makes that kind of cross-system task buildable without a tangle of fragile scripts. And because the connector is decoupled from the model, you are not locked to one vendor. If a cheaper or better model appears — and in 2026 they appear often — you can switch the brain without rebuilding the wiring. For a business watching costs, that optionality is worth real money.
Where it pays off, and where to be careful
MCP is a strong fit for some jobs and overkill for others. The table below sorts them.
| Good fit for MCP | Probably not worth it yet |
|---|---|
| agents that touch several internal systems | a single, simple one-off API call |
| tools you want reusable across many models | a throwaway prototype you will delete next week |
| letting a coding assistant read your codebase | a static FAQ chatbot with no actions |
| connecting AI to SaaS you already pay for | cases where any tool access is too risky |
The caution is real and worth stating plainly: the same survey that tracked MCP’s growth named security as the leading blocker. That makes sense. The protocol’s whole job is to give an AI controlled access to sensitive systems — your database, your inbox, your payment tools. Easier access cuts both ways. An agent that can read an email can be told, by text hidden inside that email, to do something it should not. We wrote about that class of attack in our piece on prompt injection and tool poisoning, and MCP does not remove the risk — it concentrates it at a single, well-defined boundary.
The good news is that a single boundary is exactly what you can defend. The pattern that works is to treat every MCP server as a gate, not an open door.
AI agent (client)
|
v
+-----------------------------+
| MCP server = the gate |
| - authenticate the caller |
| - allow only listed tools |
| - require approval to write |
| - log every single call |
+-----------------------------+
|
v
your real systems (DB, CRM, payments)Read actions can be fairly open; actions that change or send something — money, messages, records — should require explicit approval and leave an audit trail. Scope each server to the minimum it needs. The 2026 protocol roadmap is itself focused on exactly these enterprise concerns: SSO-integrated authentication, governance, and audit trails, because the early adopters all hit the same wall.
How to adopt it without overcommitting
You do not need a grand platform to start. Pick one painful, repetitive workflow that spans two or three systems — the kind your team does by hand every day. Wrap just those systems in small MCP servers, with read access first and writes gated behind approval. Connect one agent, measure whether it actually saves time, and watch the logs closely for a week.
If it holds up, add the next system. Because MCP is a standard, each server you add becomes reusable for the next use case, so the second project is cheaper than the first and the third cheaper still. That compounding is the real reason the standard is spreading — not hype, but the fact that the work stops being thrown away.
The pattern to remember is simple. Standardize the connection, decouple it from any one model, and put a real gate on every tool that can act. Do that, and an AI agent stops being a clever demo and becomes something you can actually trust with your business systems.
If you want help deciding which workflows are worth connecting — and building the MCP servers and guardrails to do it safely — see our software engineering services or talk to Bee Mata.