How We Built an AI Agent from Natural Language Using Our MCP

Editor's note: first published on 1 August 2025. Some product names in this post have changed since then.

At Vortex IQ, we believe the future of software is agentic, where intelligent agents transform natural language into real business actions. This blog post unpacks how we made that a reality using our proprietary Model Context Protocol (MCP) Server, enabling any user to create AI agents directly from plain English instructions.

The Vision: Natural Language → Intelligent Action

Most startups are drowning in APIs but starving for automation. Traditional RPA and scripting tools fall short, they require technical effort, can’t reason, and don’t scale well across departments.

Our goal was simple: Could a business user describe what they want, and have an AI agent automatically execute it using APIs, no code, no fuss?

The answer is yes. And it all starts with the MCP Server.

What is the MCP Server?

The Model Context Protocol (MCP) Server is our innovation layer that:

Accepts natural language input from users
Translates it into structured agent intent
Maps the intent to a set of API workflows
Executes those workflows using authenticated, secure API calls
Returns live results, summaries, or completes actions autonomously

Think of it as a bridge between what the user means and what the system does, without needing human developers in the loop.

The Build: Step-by-Step Breakdown

1. Schema Mapping for API Calls

We first created structured JSON schema files for the target API (e.g., BigCommerce, Shopify, Google Analytics), categorised into:

number_fields
text_fields
boolean_fields
date_fields

Each field includes:

A plain English description
Data type enforcement
Real examples from production

These schemas allowed our agents to reason about inputs, validate them, and auto-generate queries or payloads.

2. Prompt Compiler Engine

Next, we built a prompt compiler that transforms natural language into JSON intent objects.

Example input:

“Update the price of SKU ABC123 to £15.99 and make it visible in all channels.”

Output:

json

CopyEdit

{

“action”: “update_product”,

“sku”: “ABC123”,

“price”: “15.99”,

“visibility”: “all”

}

We use a hybrid of OpenAI models (GPT-4o) and domain-specific prompts to ensure reliability and determinism.

3. Agent Skill Builder

Each skill is a modular unit within the agent, essentially a microservice with a well-defined purpose (e.g., update stock, create discount, fetch analytics).

The MCP server connects skills to:

HTTP APIs (REST, GraphQL)
Auth credentials (OAuth, API keys)
Condition flows (e.g., “if this fails, try fallback”)

Skills are stackable, meaning agents can compose workflows like LEGO blocks.

4. Live Execution Layer

Once the intent is resolved, the MCP server routes it to the correct agent and skill set, handling:

Auth tokens securely
Real-time response handling
Error fallback logic
Post-processing summaries

Example: If an agent updates 300 SKUs, it returns a completion report, success/failure ratio, and suggested next steps.

Security & Control

We designed the entire platform to be enterprise-grade:

All API keys and credentials are encrypted at rest and in transit
Role-based access control governs who can do what
Admins can sandbox actions before pushing to production

Agents can also be deployed in staging environments, allowing teams to test AI-driven changes safely.

Real-World Use Case

A leading retail brand used our agent to:

Optimise 1,200 product descriptions for SEO
Sync prices across 3 regional stores
Create discount campaigns for underperforming SKUs

All of it was driven by conversational input, managed by our MCP server and executed via 5 chained agents.

What’s Next

We’re now building a no-code Agent Studio where:

Anyone can create agents from prompt templates
Agents can be scheduled, triggered, or embedded in workflows
Partners can monetise their own agents in our Agent Marketplace

We see a future where every team, from marketing to inventory, has their own agent, working 24/7.

How We Built an AI Agent from Natural Language Using Our MCP Server