What MCP Does
Before MCP, every AI tool integration was custom. Connecting Claude to a database required bespoke code. Connecting it to GitHub required different bespoke code. MCP provides a standard JSON-RPC protocol that any tool can implement, and any AI model can consume. Build one MCP server for your tool, and it works with Claude, Claude Code, and any MCP-compatible client.
As of June 2026, there are 3,000+ MCP servers available — covering databases, APIs, file systems, browsers, cloud services, and developer tools. Claude Code ships with built-in MCP support, making tool integration a one-line configuration.
MCP Architecture
MCP uses a client-server model. The MCP client (Claude Code, Claude Desktop, IDE extensions) discovers and calls tools. The MCP server exposes tools, resources, and prompts via JSON-RPC over stdio or HTTP transport. Each server declares its capabilities, and the client presents them to the model as available tools.
Servers can expose three primitives: tools (functions the model can call), resources (data the model can read), and prompts (pre-built prompt templates). This separation allows fine-grained access control — a server can expose read-only resources without giving write tool access.
Security Considerations
MCP servers run with the permissions of their host process. A malicious or poorly-written server can execute arbitrary code, access the filesystem, or make network requests. As of May 2026, a security audit found vulnerabilities in 66% of popular MCP servers. Always review server source code and run untrusted servers in sandboxed environments.