Skip to main content

Fiche #3: MCP Protocol — The USB-C of LLMs

MCP (Model Context Protocol) is an open standard that lets any AI agent discover and invoke any tool through a single interface — like USB-C replaced proprietary chargers, MCP replaces custom AI integrations.

5 Key Points​

1. One Protocol, Any AI Agent, Any Tool​

Before MCP, connecting Claude to your CRM required custom code. Connecting GPT to the same CRM required different custom code. MCP defines a universal interface: tools/list to discover, tools/call to invoke. Write the integration once, every AI agent can use it.

BEFORE MCP                           WITH MCP
┌────────┐ ─custom─► ┌─────┐        ┌────────┐
│ Claude │ ─custom─► │ CRM │        │ Claude │──┐
└────────┘           └─────┘        └────────┘  │
┌────────┐ ─custom─► ┌─────┐        ┌────────┐  │  MCP    ┌─────┐
│  GPT   │ ─custom─► │ CRM │        │  GPT   │──┼────────►│ CRM │
└────────┘           └─────┘        └────────┘  │         └─────┘
┌────────┐ ─custom─► ┌─────┐        ┌────────┐  │
│ Gemini │ ─custom─► │ CRM │        │ Gemini │──┘
└────────┘           └─────┘        └────────┘
 6 integrations                      1 integration

2. The Three Primitives​

MCP defines three resource types that cover the full AI-tool interaction surface:

PrimitivePurposeExample
ToolsActions the AI can invokecreate_invoice, search_orders
ResourcesData the AI can readDatabase schemas, config files
PromptsReusable prompt templates"Summarize this API response"

3. STOA = MCP Gateway for the Enterprise​

Raw MCP has no authentication, no multi-tenancy, no rate limiting. STOA adds the enterprise layer:

4. Transport: HTTP/SSE (Today), Streamable HTTP (Tomorrow)​

MCP uses Server-Sent Events over HTTP for streaming responses. This works through firewalls, load balancers, and CDNs — no WebSocket upgrades required. The protocol version 2024-11-05 is production-stable.

TransportStatusUse Case
HTTP/SSEProductionStreaming responses, real-time
Streamable HTTPEmergingSimplified bidirectional

5. Tool Registration via Kubernetes CRDs​

In STOA, tools are declared as Kubernetes Custom Resources. GitOps-compatible, version-controlled, auditable:

apiVersion: gostoa.dev/v1alpha1
kind: Tool
metadata:
  name: create-invoice
  namespace: tenant-acme
spec:
  displayName: Create Invoice
  description: Creates a new invoice in the billing system
  endpoint: https://billing.acme.com/v1/invoices
  method: POST

Apply with kubectl apply, and the tool is immediately discoverable by authorized AI agents via tools/list.

Objections & Answers​

ObjectionAnswer
"MCP is too new, it's risky"MCP was created by Anthropic and has been adopted by Claude, Cursor, Windsurf, and dozens of open-source projects. The spec is stable (2024-11-05).
"We already have REST APIs, why add MCP?"You keep your REST APIs. MCP is the protocol AI agents use to discover and call them. STOA bridges MCP to REST — your backends don't change.
"What about OpenAI function calling?"Function calling is vendor-specific. MCP is an open standard. STOA supports both — agents using function calling can still go through the gateway.
"Our APIs are internal, AI agents shouldn't access them"That's exactly what the gateway controls. OPA policies define which agents can access which tools. No policy = no access.

Further Reading​