How I Built a Production MCP Server with Kiro

Emma Xiang — Thu, 04 Dec 2025 17:16:13 +0000

How I Built a Production MCP Server with Kiro IDE

TL;DR: I built FeedMob AdPilot, a production-ready MCP server with 6 AI-powered tools, interactive UI components, and PostgreSQL persistence using Kiro IDE. This is the story of how Specs, Steering Docs, Agent Hooks, and MCP integration transformed my development workflow.

🎯 The Challenge

Creating effective advertising campaigns requires expertise across multiple domains—understanding audiences, crafting copy, designing visuals, and optimizing for platforms. I wanted to build an AI system that could handle this entire workflow conversationally.

The technical challenge? Stitching together technologies that weren't designed to work together:

FastMCP for the MCP server
@mcp-ui/server for interactive components
Claude Agent SDK for AI processing
PostgreSQL for data persistence
Next.js for the client UI

🏗️ What I Built

FeedMob AdPilot transforms natural language campaign descriptions into complete advertising assets:

Parse Requirements → Extract campaign parameters from conversation
Conduct Research → Generate comprehensive market reports
Generate Copy → Create platform-optimized variations
Generate Images → Produce visual concepts
Create Mixed Media → Combine assets into final creatives
Campaign Management → Store everything in PostgreSQL

Live Demo: https://sc80sggw884kw448kc0sckw0.coolify-dev-pa.tonob.net/

🎨 The Kiro Difference

Spec-Driven Development (10 Features)

I created formal specifications with three documents each:

.kiro/specs/generate-ad-copy/
├── requirements.md    # User stories + acceptance criteria
├── design.md          # Architecture + correctness properties
└── tasks.md           # Implementation checklist

Impact: Every line of code traces back to a requirement. Zero missed edge cases.

Steering Documents (7 Persistent Contexts)

I taught Kiro once through always-included docs:

tech.md - Technology stack and commands
fastmcp-integration.md - FastMCP patterns
mcp-ui-integration.md - UI component creation
agent-tool-patterns.md - Claude SDK integration
claude-agent-skills.md - Agent skill development
color-guidelines.md - Design system

Impact: Saved 8+ hours of repetitive context across 100+ conversations.

Agent Hooks (4 Automated Workflows)

{
  "name": "Security Pre-Commit Scanner",
  "when": { "type": "fileEdited", "patterns": ["**/*"] },
  "then": {
    "type": "askAgent",
    "prompt": "Scan for API keys, tokens, credentials..."
  }
}

Hooks I automated:

Security scanning on every file edit
MCP-UI docs checker for API changes
MCP builder guidance for server files
Agent skill builder for new tools

Impact: Prevented 3+ major bugs before they reached production.

MCP Integration (6 Servers)

Extended Kiro with specialized capabilities:

mcp-ui-docs - Always-current mcp-ui documentation
agent-skills - Skill creator and MCP builder
context7 - Instant library docs (FastMCP, Zod)
exa - Code examples and patterns
fetch - External documentation
sequential-thinking - Complex problem decomposition

Impact: The mcp-ui-docs server alone saved days of debugging outdated APIs.

💡 The Most Impressive Moment

I told Kiro:

"Create a new MCP tool for generating ad images based on the pattern in #File src/tools/generate-ad-copy.ts"

Kiro generated in ONE response:

✅ Complete Zod schema with validation
✅ Agent service with Claude SDK integration
✅ UI factory with interactive components
✅ Error handling with all edge cases
✅ Proper FastMCP registration

It understood the entire project pattern from steering docs—no repeated explanations needed.

🎭 The Frankenstein Architecture

This project is a Frankenstein's monster in the best way—incompatible technologies brought to life:

┌─────────────────────────────────────────┐
│  MCP Tools (FastMCP)                    │
│  - parseAdRequirements                  │
│  - conductAdResearch                    │
│  - generateAdCopy                       │
│  - generateAdImages                     │
└──────────────┬──────────────────────────┘
               │
┌──────────────▼──────────────────────────┐
│  Agent Services (Claude Agent SDK)      │
│  - 6 specialized agents                 │
│  - Custom skill plugins                 │
└──────────────┬──────────────────────────┘
               │
┌──────────────▼──────────────────────────┐
│  UI Factories (mcp-ui)                  │
│  - Interactive components               │
│  - Real-time updates                    │
└──────────────┬──────────────────────────┘
               │
┌──────────────▼──────────────────────────┐
│  Data Layer (PostgreSQL)                │
│  - Campaign persistence                 │
└─────────────────────────────────────────┘

🚧 Real Challenges I Faced

Gateway Timeouts: Long-running AI operations hitting limits
Tool Call Timeouts: Optimizing agent processing times
Docker Dependencies: Image generation missing libraries
Chatbot Compatibility: Existing mcp-ui clients didn't work—built custom Next.js interface
Agent Skill Design: Getting consistent JSON output from LLMs

Each challenge was solved through iterative fixes with Kiro's help.

📊 By The Numbers

🎯 6 MCP tools built and deployed
🤖 6 agent services with custom skills
📝 10 formal specs with complete traceability
📚 7 steering docs for persistent context
🪝 4 agent hooks for automated workflows
🔌 6 MCP servers integrated
⏱️ 6 days from concept to production
💬 100+ conversations with Kiro

🎓 What I Learned

1. Hybrid Vibe + Spec Wins

Spec Mode for complex features → Structure, traceability, testing
Vibe Mode for iterations → 10x faster bug fixes and adjustments

Neither alone is optimal. The combination is magic.

2. Steering Docs Are Superpowers

Teaching Kiro once vs. teaching it every conversation:

Before: 5 minutes/conversation × 100 = 500 minutes wasted
After: 0 minutes repetition = 8+ hours saved

3. MCP Integration Enables The Impossible

Without mcp-ui-docs, I would have:

Manually searched documentation
Used outdated API patterns
Spent days debugging production issues

With it: Zero breaking changes, always-current APIs.

4. Hooks Catch Issues Early

The Security Pre-Commit Scanner found issues immediately, not in code review days later.

🚀 What's Next

Immediate priorities:

Fix gateway timeout issues
Optimize agent processing times
Complete Docker image generation setup

Future features:

Video ad generation for TikTok/Reels
A/B testing integration
Multi-platform campaigns
Real-time analytics
Template library

🏁 Final Thoughts

Kiro isn't just an AI coding assistant. It's a development environment where AI understands your project deeply and assists at every stage—from requirements to deployment.

The combination of:

✅ Specs for structured planning
✅ Steering for persistent context
✅ Hooks for automation
✅ MCP for extended capabilities
✅ Vibe for natural iteration

...enabled me to build in 6 days what would have taken 2+ months traditionally.

This isn't incremental improvement. This is transformation.

🔗 Links

Live Demo: https://sc80sggw884kw448kc0sckw0.coolify-dev-pa.tonob.net/
GitHub: https://github.com/feed-mob/feedmob-adpilot-mcp
MCP Server: https://b8ggscok0gkgw848444gcokw.coolify-dev-pa.tonob.net/mcp

Built for Kiroween 2025 🎃

What's your experience with AI-assisted development? Have you tried spec-driven development with AI? Drop your thoughts in the comments! 👇

kiro #mcp #ai #devtools #hackathon #fastmcp #claude

Forem: Emma Xiang