38 servers. Over 530 tools. That sounds like enterprise infrastructure at a corporation. It's the AI infrastructure of a web design agency in Mallorca. How it got there, why MCP is the critical standard, and what these servers actually do -- that's what this article is about.
What Is MCP?
Model Context Protocol (MCP) is an open standard by Anthropic that gives AI models access to external tools. Instead of telling a chatbot "here's some text, do something with it," an MCP-enabled agent can directly access databases, send emails, manage files, or process payments.
The comparison: REST APIs connect software to software. MCP connects AI to software. The difference is fundamental -- MCP servers describe not just their endpoints but also when and why a tool should be used. The AI model decides on its own which tool it needs for a task.
Our MCP Architecture
We operate our MCP servers in multiple layers:
Global Servers (12 Servers -- Always Active)
These servers load in every session, regardless of the project:
- Memory -- Our central memory system with 50 tools. Sessions, decisions, learnings, knowledge graph.
- Code Intelligence -- Code analysis, dependency tracking, architecture drift detection. 69 specialized tools for code quality.
- Web Research -- Search engine integration, deep research, news research.
- GitHub -- Repository management, code search, pull request management.
- Library Docs -- Live documentation for frameworks and libraries.
Project Servers (19 Servers -- Project-Specific)
These servers load only in StudioMeyer sessions and cover operational business:
Content & Communication:
- Social Media -- 14 layout templates, 7 formats, publishing to LinkedIn and Instagram. Generates carousels, stories, and post images with brand kit.
- Video -- Recording, editing, text-to-speech. Produces screencasts and explainer videos.
- PDF -- 15 templates, trilingual (DE/EN/ES), brand kit integration. From proposals to invoices.
- Email -- OAuth2 integration for Gmail and Outlook, templates, scheduled sending, five mailboxes.
- Media -- Image generation with three providers, smart routing, five style presets.
Business Processes:
- CRM -- Contact management, interaction tracking, pipeline management.
- Analytics -- Website statistics, page views, referrers, real-time data.
- Payments -- Stripe integration with checkout, subscriptions, and invoices.
- Calendar -- Google Calendar with recurring events, Meet links, and conflict detection.
- Scheduling -- Job management with PostgreSQL persistence, cron, and webhooks.
Infrastructure:
- Auth -- JWT, API keys, role-based access control, OAuth, sessions.
- Storage -- S3/R2 cloud storage for files and assets.
- Database -- PostgreSQL management: queries, schema management, migrations.
- Forms -- Form management with validation, webhooks, and export.
- Notifications -- Email, SMS, push, webhooks, and in-app notifications.
Cross-Project Intelligence
Additionally, we operate specialized servers that work across projects:
- Cross-Repo Intelligence -- Analyzes dependencies between 18 projects. Detects when a change in Project A impacts Project B.
- Fleet Health Monitor -- Monitors the health status of all MCP servers. Health checks, config audits, tool inventory.
- Architecture Tracker -- Detects architectural drift through snapshot comparisons. Warns when the codebase evolves in an unintended direction.
Security
With over 530 tools, security isn't optional -- it's mandatory. Our measures:
At the Server Level:
- Each agent has exclusive MCP servers. No agent shares tools with another.
- Tool limits: 10 to 20 tools per agent (research shows accuracy drops beyond 30 tools).
- Server instructions prevent dangerous operations ("no delete," "never guess IDs").
At the Code Level:
- Four security fixes for critical vulnerabilities: Command Injection (Scheduler), JWT Secret Enforcement (Auth), Path Traversal (Storage), ReDoS (Forms).
- SQL injection prevention: raw queries replaced with parameterized queries.
- SSRF protection: private IPs and non-standard ports are blocked.
At the Infrastructure Level:
- All servers run in Docker containers with memory limits.
- Dual transport: stdio for local use, Streamable HTTP for remote access.
- OAuth 2.1 for production HTTP servers.
From 38 to 15: Consolidation
38 servers are too many. The insight came from practice: with 530 tools in a session, the quality of the AI model's tool selection drops. The sweet spot is 30 to 50 tools per session.
Our consolidation plan:
- Research + Tavily → one unified research server
- Bookkeeping + Billing → one finance server
- Inbox + Outreach + Notifications → one communications server
- Calendar + Scheduler → one time management server
- Generate + Effects + Media → one content server
- Website + SEO → one web analysis server
The goal: 12 to 15 servers with an MCP gateway in front handling tool filtering, usage tracking, and OAuth auth.
Open Source
Six of our MCP servers are open source and freely available:
- mcp-video -- Video production (recording, editing, TTS)
- agent-fleet -- Agent orchestration
- ai-shield -- Security for AI chatbots
- darwin-agents -- Self-improving agents
- email-mcp -- Email integration
- mcp-searxng -- Search engine integration
Together, these repos have recorded over 340 clones in 14 days from 180 unique users. This shows: the demand for specialized MCP servers is real.
What We've Learned
MCP Is Not a REST Replacement
MCP and REST solve different problems. REST connects deterministic software systems. MCP gives AI models the ability to use tools intelligently. Both coexist -- our MCP servers often use REST APIs internally.
Tool Descriptions Matter More Than Tool Implementation
An MCP server is only as good as its descriptions. If the AI model doesn't understand when a tool should be used, it gets used incorrectly or not at all. We invest more time in server instructions and tool descriptions than in the actual implementation.
Fewer Tools, Better Results
58 tools in a single server (our old Generate server) were too many. After consolidating to 10 to 12 tools per server, the hit rate for tool selection improved significantly. Research confirms: beyond 30 tools per context, accuracy drops significantly.
Conclusion
Building an MCP infrastructure doesn't mean running as many servers as possible. It means making the right tools available at the right time in the right context. Less is more -- but the few must be excellently described, securely implemented, and intelligently filtered.
