Lab 1: How AI Connects to Your Systems (MCP)

What You'll Learn

In this lab, you'll see how AI connects to external data sources directly — instead of relying on data you paste into the chat. That's what MCP (Model Context Protocol) does.

📋

Without MCP

You paste data into the chat, or the AI reads files from your workspace. Manual, one-at-a-time.

🔌

With MCP

AI connects to websites, databases, APIs, and legislation directly. Ask in plain English — it fetches the answer.

💡 What is MCP?

Model Context Protocol (MCP) is an open standard that connects AI to external tools and data sources. Think of it as a USB-C port for AI:

MCP Server = the data source (website, database, API, legislation database)
MCP Client = Kiro (the AI that uses the data)
You don't write code — you configure a connection, and the AI can query it

Today you'll see two MCP servers in action: one that fetches web pages, and one that searches AWS documentation.

Part	Duration	What you'll see
Part A	10 min	Activate Fetch MCP → AI reads AnyCompany's policy page and answers questions about coverage
Part B	12 min	Configure AWS Docs MCP → AI searches AWS documentation and answers architecture questions
Part C	5 min	Vision — what MCP servers would AnyCompany need in production?

Part A: Fetch MCP — AI Reads the Web

Kiro comes with a Fetch MCP server pre-configured — it lets your AI agent read any public web page and return its content. But first, we need to make sure it's enabled.

Step 0: Activate the Fetch MCP Server

The Fetch MCP server is pre-installed but may be disabled in your environment. Let's ask Kiro to enable it — no manual editing needed:

PROMPT — Copy & paste into Kiro

Enable the Fetch MCP server in my user MCP configuration. Change "disabled" from true to false for the fetch server, then save the file.

Kiro will ask permission to edit the config file. Click "Allow" when you see the prompt — then it makes the change and saves automatically.

✅ Checkpoint: Look at the MCP SERVERS section in the left sidebar (bottom of the Kiro panel). "Fetch" should change from Disabled to Running within a few seconds.

💡 What just happened?

You asked the AI to configure its own connection to an external tool. Behind the scenes, Kiro edited a small config file (mcp.json) that lists which MCP servers to connect to. This is the same pattern for any MCP server — add it to the config, and the AI gains new capabilities. No coding required.

Scenario: A customer asks about critical illness coverage

Imagine a customer service agent needs to quickly check what's covered under AnyCompany's Critical Illness plan. Instead of searching through internal documents, the AI reads the product page directly:

PROMPT — Copy & paste into Kiro

Fetch the AnyCompany Insurance product guide from this URL: https://sing.lab.mywcloud.net/anycompany-policies.html Then answer these questions: 1. How many critical illnesses are covered under AnyCare Critical? 2. What is the maximum sum assured available? 3. If a customer has cancer and recovers, can they claim again if it comes back? 4. What is the survival period requirement? Format your answers as a clear summary a customer service agent could read to a caller.

⚠️ Important: When Kiro fetches the page, look at the tool call in the chat. You should see it using the mcp_fetch_fetch tool (the MCP server you just enabled). If you see web_fetch instead, the MCP server may not have connected yet — go back and check the MCP SERVERS panel shows "Fetch" as Running.

✅ What to observe:

Kiro fetches the web page content automatically
It extracts specific answers from the product information
The answers are accurate — grounded in the actual page content, not hallucinated
It formats the response for the intended audience (customer service agent)

Try another query — comparing products:

PROMPT — Product comparison

Based on the AnyCompany product guide you just read, create a comparison table for a customer who is deciding between the three AnyCare Shield tiers (Basic, Enhanced, Premier). Focus on: annual claim limit, ward class, daily room & board, and annual premium for a 30-year-old. Add a recommendation at the bottom: which tier suits (a) a budget-conscious young professional, (b) a family breadwinner, and (c) a high-net-worth individual?

💡 Why this matters for AnyCompany:

Today, your customer service team searches through PDF product brochures or internal wikis to answer questions. With Fetch MCP, an AI agent can read your product pages, policy documents, or even competitor websites — and synthesize answers in seconds. This is the foundation for an intelligent customer service agent.

Part B: AWS Documentation MCP — AI Searches a Knowledge Base

Now let's connect a more powerful MCP server — one that searches across the entire AWS documentation library and returns specific, up-to-date answers about cloud services.

📖 About this MCP server

The AWS Documentation MCP Server is an open-source server maintained by AWS Labs that provides:

Search across all AWS documentation (thousands of pages)
Read specific documentation pages and extract sections
Recommend related pages based on what you're reading
No API key required — runs locally via uvx (same as Fetch)

Step 1: Configure the MCP connection

Just like Part A, we'll ask Kiro to add a new MCP server. Paste this into the chat:

PROMPT — Add MCP server

Add a new MCP server to my workspace MCP configuration (.kiro/settings/mcp.json). The server details are: Name: "aws-docs" Command: uvx Args: ["awslabs.aws-documentation-mcp-server@latest"] Env: {"FASTMCP_LOG_LEVEL": "ERROR"} Add it to the workspace config file and make sure disabled is set to false.

Kiro will ask permission to edit the config file. Click "Allow" — then wait 15–20 seconds for the server to download and start (first time only).

⏳ First-time setup

The AWS Documentation MCP server downloads its package the first time (~15–20 seconds). If you see Connection Failed in the MCP panel, click "Retry" — it sometimes needs a second attempt after the initial download completes.

✅ Checkpoint: In the MCP SERVERS section (left sidebar), you should see "aws-docs" with a Connected status and its tools listed (search_documentation, read_documentation, recommend).

💡 Same pattern, different data source

Notice: both Fetch and AWS Docs use the same setup pattern — uvx command, a package name, and that's it. The AI now has two MCP tools: one reads web pages, the other searches AWS documentation. Same protocol, different capabilities.

Step 2: Query AWS documentation — insurance-relevant questions

Now ask questions about AWS services that AnyCompany might use — in plain English:

PROMPT — AI services for insurance

Using the AWS Documentation MCP server, search for information about: 1. What is Amazon Bedrock and what foundation models does it support? 2. How does Amazon Bedrock Guardrails work to filter harmful content? 3. What is Amazon Bedrock Agents and how can it automate multi-step tasks? For each answer, give me a plain-English summary that a business leader at an insurance company would understand. Focus on what it does, not how to code it.

PROMPT — Data protection & compliance

Search AWS documentation for: 1. How does AWS handle data residency — can we keep data in Singapore? 2. What security certifications does AWS have that matter for financial services? 3. How does Amazon Bedrock handle data privacy — does AWS use our data to train models? Summarize each answer in 2-3 sentences. This is for a VP of Security evaluating whether to use AWS AI services for insurance operations.

PROMPT — Architecture scenario

AnyCompany Insurance wants to build an AI-powered claims processing system. Search AWS documentation to help me understand: 1. Which AWS service would we use to extract text from medical reports and claim forms? (hint: look for document processing) 2. Which service would we use to build the AI agent that orchestrates the claims workflow? 3. How would we store and search through historical claims data for the AI to reference? For each, give me the service name, what it does, and why it fits our claims use case.

🔍 What to observe:

The AI searches across AWS documentation — not using general knowledge
It returns specific, current information from official docs
It can cross-reference multiple services to answer architecture questions
The answers are grounded in documentation, not hallucinated

💡 Compare this to today's workflow:

	Current approach	With MCP
Research	Google → open 10 tabs → read each page → take notes	"What AWS services handle document processing?"
Compare	Open multiple doc pages, manually compare features	"Compare Textract vs Comprehend for medical reports"
Architecture	Read whitepapers, watch re:Invent talks, ask a solutions architect	"What services would I need for an AI claims system?"
Time	Hours of research per question	30 seconds

🏢 For AnyCompany: imagine this with YOUR documentation

You just saw the AI search AWS docs. Now imagine the same pattern connected to:

AnyCompany's internal wiki — "What's our claims escalation process?"
Policy product guides — "What's covered under AnyCare Critical?"
MAS regulatory circulars — "What are the latest Fair Dealing requirements?"
HR policies — "What's our remote work policy for the Singapore office?"

Same MCP protocol, same setup pattern — just pointed at different data sources. That's the power of a standard protocol.

⚡ Bonus (Optional): From Reading Docs to Taking Action

So far, your AI has been reading — fetching web pages and searching documentation. But MCP can also let AI interact with live systems. Let's add one more MCP server that can actually run AWS commands:

PROMPT — Add AWS API MCP server

Add a new MCP server to my workspace MCP configuration (.kiro/settings/mcp.json). The server details are: Name: "aws-api" Command: uvx Args: ["awslabs.aws-api-mcp-server@latest"] Env: {"AWS_REGION": "us-east-1"} Add it to the workspace config file and make sure disabled is set to false.

Click "Allow" when prompted. Wait 15–20 seconds for the first-time download.

✅ Checkpoint: In the MCP SERVERS section, you should see "aws-api" with a Connected status.

Now let's use it — ask the AI to query your actual AWS account:

PROMPT — AWS API interaction

Using the AWS API MCP server: 1. List the S3 buckets in this AWS account 2. List the Bedrock foundation models available in us-east-1 3. Check what region this account is configured for Summarize what you find — I want to understand what AWS resources are available to us.

🔍 What to observe:

The AI makes real AWS API calls — not searching docs, but querying live infrastructure
It returns actual bucket names, model IDs, and account details from your environment
This is the same pattern that would let an agent check claims database status, query policy systems, or monitor infrastructure health

💡 The progression you just experienced:

Part	What the AI did	MCP capability
A	Read a web page	Fetch (read external content)
B	Searched a knowledge base	AWS Docs (search + retrieve)
Bonus	Queried live infrastructure	AWS API (read + act)

This is the spectrum of MCP: from passive reading → active searching → live system interaction. Each step gives the AI more capability — and each requires appropriate guardrails.

Part C: What MCP Servers Would AnyCompany Need?

You've seen two MCP servers in action. In production, AnyCompany would connect AI agents to many more data sources — all through the same MCP protocol.

🏗️ AnyCompany's MCP Architecture (Vision)

MCP Server	What it connects to	Example queries
Policy Admin	Core policy system (Oracle/SAP)	"Show all policies lapsing this month" · "What's the status of POL-2024-8847?"
Claims System	Claims management platform	"List all pending claims over $50K" · "Average processing time for critical illness claims?"
Customer 360	CRM + interaction history	"Full profile for customer CUST-88421" · "Customers with 3+ policies not contacted in 6 months"
Singapore Law	Legislation database (SSO)	"Insurance Act requirements for policy illustrations" · "PDPA consent for health data"
AWS Documentation	Cloud services knowledge base	"What AWS service handles document extraction?" · "Bedrock data privacy guarantees"
MAS Circulars	Regulatory notices & guidelines	"Latest MAS notice on technology risk management" · "Fair Dealing outcome requirements"
Actuarial Models	Pricing & reserving systems	"Current mortality rates for age 45 male non-smoker" · "IFRS 17 reserve for product X"
Document Store	Policy documents, medical reports	"Retrieve the medical report for claim CLM-2024-0847" · "Find the policy contract for POL-123"
Web / Fetch	Public websites, competitor info	"Read the latest MAS media release" · "Compare our CI coverage with competitor X"

Each MCP server is built and maintained by the team that owns that data source. The AI agent decides which server to query based on the question — just like a human would know which system to check.

🎯 For the Agent Design Canvas (Lab 3)

When you fill in the "Data Sources" section of your Agent Design Canvas, think about which MCP servers your agent would need. The pattern is always the same:

What data does the agent need to do its job?
Where does that data live today? (database, API, website, document)
That becomes an MCP server connection

🌐 MCP is growing fast

The MCP ecosystem already has hundreds of servers for different data sources:

Databases: PostgreSQL, MySQL, SQLite, MongoDB
APIs: Slack, Jira, GitHub, Salesforce, ServiceNow
Documents: Google Drive, SharePoint, Confluence
Specialized: Singapore Law, AWS Documentation, financial data feeds

Your tech team doesn't need to build everything from scratch — many MCP servers already exist. For custom internal systems, building an MCP server is typically a few hundred lines of code.

📋 Summary: The MCP Stack

Layer	What it does	You saw this in
Steering	Global rules (currency, PII, ratings)	Lab 2, Step 1
Skills	On-demand expertise (risk assessment, compliance)	Lab 2, Steps 2-5
Hooks	Auto-triggers (new file → run skill)	Lab 2, Step 3
MCP	Data connections (web, legislation, databases)	This lab

🎉 What You've Seen

Fetch MCP — AI reading a web page and extracting structured answers about insurance products
AWS Documentation MCP — AI searching thousands of documentation pages and returning specific, current answers
The MCP pattern — same protocol, different data sources, AI decides which to query
Production vision — how AnyCompany would connect AI to policy admin, claims, CRM, and regulatory systems

💡 Key takeaway for your team

MCP turns AI from a "chatbot that knows general things" into a "system that can query YOUR data." The protocol is open, the ecosystem is growing, and the setup is configuration — not custom development. When you design your agent in Lab 3, think about what data sources it needs. Each one becomes an MCP connection.

← Workshop Home Lab 2: Skills →

🔌 Lab 1: How AI Connects to Your Systems

What You'll Learn

Without MCP

With MCP

Part A: Fetch MCP — AI Reads the Web

Step 0: Activate the Fetch MCP Server

Scenario: A customer asks about critical illness coverage

Try another query — comparing products:

Part B: AWS Documentation MCP — AI Searches a Knowledge Base

Step 1: Configure the MCP connection

Step 2: Query AWS documentation — insurance-relevant questions

⚡ Bonus (Optional): From Reading Docs to Taking Action

Part C: What MCP Servers Would AnyCompany Need?

🏗️ AnyCompany's MCP Architecture (Vision)

📋 Summary: The MCP Stack

🎉 What You've Seen