MCP explained

MCP vs API: How Model Context Protocol Works With the APIs You Already Have

A plain-English comparison for operators deciding how Claude and other AI tools should reach their systems, with the security controls that matter and when a custom MCP server is worth building.

By Chad Westleigh · Updated October 16, 2026

Book a scoping call See MCP integration services

MCP vs API: the short answer

An API (application programming interface) is the contract a system publishes so other software can read and change its data: Shopify's Admin API, HubSpot's CRM API, the Gmail API. A developer reads the documentation and writes code that calls specific endpoints in a specific order.

MCP sits one layer up. An MCP server describes a set of tools, such as look up an order or update a product price, in a form an AI model can read at runtime, then turns each tool call into the underlying API requests. The model decides which tool to use and when; the API still does the work. Our AI integrations hub shows where we use each route.

So the useful question isn't MCP or API. It's whether your code or an AI model is choosing the next step. If code is choosing, call the API. If a model is choosing, as in the custom AI agents we build, give it tools through MCP.

Model Context Protocol in one paragraph

Model Context Protocol is an open standard for connecting AI applications to external tools and data. Anthropic introduced it on November 25, 2024. In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, and reported more than 10,000 active public MCP servers, with adoption in products including ChatGPT, Cursor, Gemini and Microsoft Copilot.

The specification (version 2026-07-28 as of October 2026) defines three roles: a host, the AI application such as Claude; clients, the connectors inside the host; and servers, the services that expose capabilities. Messages use JSON-RPC 2.0. Servers can offer tools (functions the model can execute), resources (data and context) and prompts (templated workflows). Remote servers use OAuth-based authorization; local servers take credentials from their environment. For a shorter definition, see our glossary entry on Model Context Protocol.

MCP vs API comparison table

Question	API	MCP
Who calls it	Code a developer writes	An AI application, when a model decides to
How capabilities are found	Developers read documentation ahead of time	The client asks the server at runtime which tools exist, with descriptions and input schemas
Interface	Each vendor's own endpoints, formats and auth	One protocol (JSON-RPC 2.0) across every server
Authentication	Vendor-specific API keys or OAuth	OAuth 2.1-based for remote servers; environment credentials for local servers
What it replaces	Nothing: it is the system's front door	Custom glue code between each AI app and each tool
Predictability	Same input, same calls	The model may pick different tools or orders, so limits and approvals matter
Suited to	Fixed, high-volume, deterministic steps	Varied requests where a model must choose the action
Example	A workflow posts each new order to your ERP	Claude fixes a product page issue someone described in Slack

How MCP and APIs work together: the ROAM Luggage example

The agentic Shopify workflow Operator Intelligence built for ROAM Luggage shows the layers. The ROAM team requests site changes in Slack or email. A custom n8n workflow saves each request to a task repository and AI prioritizes the queue. Claude then executes each change via MCP inside a Shopify development theme, which is where a model has to decide what to edit. A ROAM team member reviews and publishes.

Deterministic steps run on workflows and APIs, the judgment step gets MCP tools, and a person holds the publish button. Critical site updates went from days to minutes and ongoing website management costs fell by more than 90%. Read the ROAM Luggage case study or see our agentic Shopify management service.

When to use the API, an existing MCP server or a custom MCP server

Call the API directly

The steps are fixed and high volume, like syncing orders or posting a weekly report. Code or a workflow tool is cheaper, faster and easier to test than a model choosing tools.

Use an existing MCP server

The vendor ships one that covers your use case. Shopify, Klaviyo and Zapier all publish MCP servers, and Klaviyo's remote server can be connected with write tools disabled.

Build a custom MCP server

No server exists, the vendor's server exposes too much or too little, or your tools need to combine several systems under your own rules, logging and approvals.

Learn more

Use neither yet

The task happens a few times a month or the process isn't written down. Fix the process first; a well-configured AI workspace may be enough.

What goes into a production custom MCP server

A custom MCP server is a small product. These are the parts we treat as non-negotiable.

Narrow tools named for business actions, such as create draft discount, not raw endpoints
Read and write tools separated, with write tools off until reads prove reliable
Least-privilege credentials, and only tokens issued for that server
Hard limits enforced in the server, such as price ranges and record counts, not just in the prompt
A log of every tool call, input and result, reviewed by a named owner
A sandbox or approval step, such as a development theme, before changes go live

MCP security: the risks that matter

MCP security deserves its own review because MCP gives a model a path to act. The specification says tools represent arbitrary code execution, that tool descriptions from untrusted servers should be treated as untrusted, and that hosts must get user consent before invoking a tool. The risks we plan for:

Prompt injection through content. An email, web page or product review returned by a tool can carry instructions the model might follow.
Untrusted servers. A third-party server's descriptions or code can be malicious, and local servers run with your user's privileges.
Token passthrough. The spec forbids MCP servers from accepting tokens that weren't issued to them, because forwarding them breaks controls and audit trails.
Confused deputy. Proxy servers in front of third-party APIs must collect per-client consent, or an attacker can gain access without the user's approval.
Over-broad scopes. One token with every permission turns a small leak into a large one; the spec recommends minimal scopes with step-up for privileged actions.

Our controls follow from these: least privilege, separated read and write tools, human approval for consequential actions, full logging and sandboxes first. Our MCP integrations service applies them to every server we build or connect.

MCP in production

Case study · DTC luggage

ROAM Luggage Case Study: Agentic Shopify Management With Claude and MCP

Minutesfor critical site updates, down from days

Read the case study

Frequently asked questions

Does MCP replace APIs?

No. Most MCP servers call an API underneath. MCP standardizes how an AI application discovers and calls those capabilities, so every AI app doesn't need custom integration code for every system. HubSpot is a good example: our n8n HubSpot integration page compares n8n's HubSpot node, Claude's HubSpot connector and the HubSpot MCP server.

Is MCP secure?

MCP is a protocol, so security depends on the implementation. The specification sets OAuth-based authorization for remote servers, forbids token passthrough and recommends minimal scopes, but you still need trusted servers, least-privilege credentials, logging and human approval for risky actions.

Who maintains the MCP standard?

Anthropic introduced MCP in November 2024 and donated it to the Agentic AI Foundation, under the Linux Foundation, in December 2025. Anthropic said MCP's governance model would remain unchanged, with maintainers making decisions in the open.

Which tools support MCP?

As of October 2026, Claude, ChatGPT, Gemini, Microsoft Copilot and Cursor support MCP in some form, as do automation platforms including n8n, Zapier and Make. Plan tiers and admin settings decide who can add custom servers. Our n8n vs Zapier vs Make comparison covers the automation side.

How long does a custom MCP server take to build?

A narrow server with a few read-only tools is a small build. Write tools with approvals, logging and testing against real cases take longer, and the API's quality and auth model matter. We scope each server as a discrete project after discovery.

Connect your AI tools to your systems safely

Tell us which systems Claude or your other AI tools need to reach. We'll tell you whether an existing MCP server, a custom one or a plain API workflow fits.

Book a scoping call

Sources

Get the Operator Intelligence briefing

A short, practical note every two weeks on what is working with AI in real businesses. No spam; unsubscribe anytime.