I’m building a system with 1 main agent (I will have more in the future) + 4 MCP servers (I will have more in the future) and would love feedback on my approach. My containerization experience is limited, so I want to make sure I’m doing this right.
My Philosophy
I want MCPs to be HTTP-reachable from outside - not just from my main agent. The idea: if I solve a problem once (like a data integration), I don’t want to copy-paste code everywhere. I want any external tool, agent, or future protocol (agent-to-agent, ag-ui, etc.) to just reach the MCP directly. That’s why I chose streamable-http transport with exposed ports.
Current Setup
Single shared Dockerfile for all services (Python 3.13, FastMCP)
Docker Compose orchestrating 5 services
streamable-http transport with exposed ports (8001:8001, etc.)
Docker internal DNS for service-to-service (http://mcp-1:8001/mcp)
Health checks + restart: unless-stopped
Concerns Raised
A teammate suggested:
- Shared Dockerfile = slow deploys when changing one MCP
- Separate Dockerfiles per MCP would be better
- Hard-mapped ports might be problematic
My Questions
Is Docker Compose + shared Dockerfile reasonable for 4-12 HTTP-reachable MCPs?
For my use case (MCPs reachable externally), is port mapping the right pattern?
At what scale should I consider separate images or different architecture?