Building a Side Project with 7 AI Agents: Lessons Learned

The Project

mcphello.com — an MCP tool directory with 900+ tools. Built and maintained entirely by AI agents.

The Agent Team

1. Collector — finds new MCP repos on GitHub
2. Analyzer — scores each tool
3. Writer — generates review articles
4. Publisher — updates the website
5. Ops Brain — coordinates and strategizes
6. Social Agent — publishes to social media
7. Dev.to Agent — writes and publishes articles

Lessons

What Worked

• Clear agent responsibilities (single purpose each)
• JSON-based data sharing between agents
• Cron-based orchestration (simple and reliable)

What Did Not Work

• Complex agent-to-agent communication (over-engineered)
• LLM-based orchestration (unreliable, expensive)
• Real-time agent networks (unnecessary for this use case)

The Stack

• Next.js 16 + Vercel (frontend)
• Python agents (backend automation)
• GitHub API (data source)
• Playwright (social media publishing)

Result

A fully automated directory that updates itself. Zero manual work after setup.

Check it out: mcphello.com

---

Any questions about the agent architecture? Happy to share details!