Model Context Protocol (MCP) Servers
MCP on the wire: hosts, clients and servers over JSON-RPC 2.0, per-request versions and server/discover, tools, resources and prompts, stdio and Streamable HTTP, OAuth for remote servers, tool poisoning, versioning and tracing.
Overview
Why this topic
An agent is only as useful as the systems it can reach. Before MCP, every AI application wrote its own adapter for every data source: N applications times M systems. The Model Context Protocol turns that into N plus M. A system exposes one MCP server, and any MCP-capable application (a chat app, an IDE, a coding agent, your own orchestrator) can use it.
This topic is about the protocol itself and running servers in production: the messages on the wire, versions and discovery, the transports, how remote servers authorize callers, and what goes wrong.
Interview questions
- Walk through exactly what crosses the wire when a host connects to an MCP server and the model calls one tool.
- stdio or Streamable HTTP: how do you choose, and what changes operationally?
- How does authorization work for a remote MCP server, and what mistakes do teams make?
- A third-party MCP server is installed in your agent. How can it attack you, and how do you defend?
- How do you version an MCP server and its tools without breaking the agents that use them?
- What do you log and measure on an MCP server in production?
Sign up free to read the full topic
More Agentic AI topics
- Evaluating Agents: Task Suites, Trajectories & Regression Gates
- Evaluating RAG & Catching Hallucinations
- Fine-tuning LLMs: SFT, LoRA/QLoRA & Preference Tuning
- Guardrails & Prompt Injection Defense
- LLM Cost & Latency Engineering
- LLM Observability: Traces, Spans & OpenTelemetry GenAI
- Multi-Agent Systems with LangGraph
- RAG in Production: Ingestion to Answer
- Real-time Voice Agents: STT → LLM → TTS
Start free on Cariara · Coding problems · System design studio