**What is an MCP Server?** An MCP server is a component within the Model Context Protocol architecture that acts as a data source or tool provider. It exposes specific capabilities—such as resources, prompts, and tools—to clients (AI hosts). Technically, it communicates using JSON-RPC 2.0 messages over either stdio or Streamable HTTP transports. The protocol is stateless, meaning the server does not infer context from previous requests; every request must contain all necessary information. Servers must also ensure user consent before executing tools or exposing sensitive data.
**Why Do Agents Need One?** Agents (AI hosts/clients) require an MCP server to securely connect to external data sources, tools, and workflows. The server provides the standardized interface that allows the agent to: 1. **Discover Tools:** Use the `tools/list` method to find available capabilities. 2. **Execute Actions:** Use the `tools/call` method to invoke specific tools, passing inputs validated against JSON Schema. 3. **Negotiate Capabilities:** Engage in capability negotiation where the server exposes what it can do, and the client may offer features like sampling.
Without an MCP server, an agent cannot standardize its communication with external systems to perform actions or access data securely and consistently.
This answer was worked out by Liminality — Physea's engine that decomposes a request, grounds each part to a real tool, and returns a reusable, checkable route.
Connect it over MCP: https://mcp.physea.ai/mcp · physea.ai