MCP and the King's Hand

With LLMs, if "Context is King," then MCP is the King's Hand—the one that actually gets shit done.

An LLM without tools is like a brain in a vat - tons of clever thoughts, but can't scratch an itch for shit. The Model Context Protocol is the pair of arms that finally lets it reach out, poke the world.

Released by Anthropic in November 2024, MCP has become the "USB-C for AI agents." Major players like Microsoft, OpenAI, and Amazon have integrated support. The official repository lists 28 clients and over 200 servers spanning everything from Google Workspace to enterprise databases.

Instead of building custom integrations for every AI use case, you write one MCP server and any AI system can use it.

The OODA Loop of Intelligence

One way to think about MCP vs. LLM is to think about where each fits within the OODA loop:

1. Observe: Gather information from the environment
2. Orient: Make sense of that information
3. Decide: Choose what to do
4. Act: Execute the decision

LLMs excel at Orient and Decide—taking messy information and figuring out what it means, then deciding what should happen next. But they need tools to Observe (read files, call APIs) and Act (write files, send messages).

MCP provides those tools.

Why Developers Should Care

MCP solves the N×M integration problem. Instead of building custom connections between every AI system and every tool, you build N + M (N AI clients + M MCP servers).

Here's what makes it work:

• Simple protocol: JSON-RPC over stdio, HTTP, or WebSocket
• Tool composability: Tools work together without knowing about each other
• Network effects: Major platforms adopting it creates momentum
• Zero marginal cost: Once you have an MCP server, adding AI workflows is free

What This Looks Like

A user asks: "Analyze our Q4 sales data and share insights with the team."

The AI agent:

1. Uses a file tool to read the sales CSV
2. Uses a Python tool to run analysis and generate charts
3. Uses a Slack tool to post results to #sales

Each tool is simple. The combination is powerful. And you didn't write any orchestration logic—the AI figured it out.

The Growing Ecosystem

Developers are building MCP servers for everything:

• Databases: Postgres, MySQL, SQLite
• APIs: GitHub, Jira, Salesforce
• File systems: Local files, S3, Google Drive
• Communication: Slack, Discord, email
• Development: Docker, Kubernetes, CI/CD

Each new MCP server makes every AI agent more capable. Classic network effects.

Current Challenges

Context management: As agents access more tools, context windows fill up fast.

Permissions: Users don't want to approve every tool call, but they don't want agents running wild either.

Error handling: When agents chain multiple tools, failures can cascade.

Tool discovery: How does an agent know which tools are available and when to use them?

Early adopters are solving these problems right now. The solutions become best practices for everyone.

What's Next

We're still early—like HTTP in 1995. The interesting stuff happens as the ecosystem matures:

• Better orchestration: Smarter tool selection and error recovery
• Specialized tools: Domain-specific servers for finance, healthcare, etc.
• Tool marketplaces: Third-party repositories with ratings and analytics
• Advanced workflows: Multi-step processes with state management

The constraint stops being "what can we build" and starts being "what should we build."

The Real Impact

MCP is infrastructure for a world where describing what you want is enough to make it happen. For developers, this means building tools that work with every AI system instead of just one.

The King's Hand doesn't rule. But it's what turns intentions into reality.

---

Good tools make good agents. Good protocols make good tools possible.