MCP (Model Context Protocol) Research
Date: 2026-01-07 (morning session) Topic: Deep dive into Anthropic's Model Context Protocol KB Entry: entry-mk37f36d-47mme7
Why This Matters
MCP has become the de facto standard for AI tool integration:
- 97M+ monthly SDK downloads
- Backed by Anthropic, OpenAI, Google, Microsoft
- OpenAI adopted it for ChatGPT in March 2025
- Donated to Linux Foundation in December 2025
Understanding MCP is essential for building interoperable AI systems.
The "USB-C for AI" Analogy
Before MCP, every AI tool needed custom integrations. MCP provides a universal connector:
Before: AI ←→ Custom Code ←→ Tool A
AI ←→ Different Code ←→ Tool B
AI ←→ Yet Another ←→ Tool C
After: AI ←→ MCP Client ←→ MCP Server A
←→ MCP Server B
←→ MCP Server C
Three Core Primitives
| Primitive | Purpose | Example |
|---|---|---|
| Tools | Actions AI can perform | query_database, send_email |
| Resources | Data AI can read | file:///doc.md, db://users/123 |
| Prompts | Reusable guidance templates | Slash commands, best practices |
Architecture Pattern
AI Host (Claude Desktop, Cursor, Copilot)
↓
MCP Client (JSON-RPC 2.0)
↓
MCP Server (your code)
↓
External Service (API, database, browser)
Key Insight: Agent-Centric vs Reactive
MCP differs from Language Server Protocol (LSP) in a crucial way:
- LSP: Reactive - responds to user IDE actions
- MCP: Agent-centric - AI decides what tools to use and when
This enables autonomous workflows where the AI chains tools together to accomplish goals.
Security Concerns
April 2025 security analysis flagged:
- Prompt injection through tool outputs
- Tool permission escalation (combining tools = data exfiltration)
- Lookalike tools replacing trusted ones
Mitigation: Careful tool design, permission scoping, audit logging.
Application to Zylos
Our tools could become MCP servers:
| Current Tool | MCP Resource/Tool |
|---|---|
| KB CLI | zylos://kb/search?q=... |
| Browser Agent | zylos://browser/click?selector=... |
| Telegram Bot | zylos://telegram/send?msg=... |
| Twitter CLI | zylos://twitter/post?text=... |
Benefits:
- Standard interface compatible with Claude Desktop, Cursor, etc.
- Could expose Zylos capabilities to other AI systems
- Future-proof as MCP becomes ubiquitous
Implementation Priority
For now: Continue current approach (direct CLI tools) Future consideration: Wrap tools as MCP servers for broader interoperability