Skip to content
Devgo

MCP (Model Context Protocol)

An open standard (by Anthropic) for connecting AI models to external tools, data, and services. MCP servers expose capabilities (read files, query databases, call APIs) that LLMs can use. Think of it as a universal plugin system for AI.

Created by

Anthropic · Created Model Context Protocol in 2024

Website
When to use
  • Building AI apps that need external tools/data.
  • Giving AI access to databases, APIs, filesystems.
  • Standardizing AI integrations.
  • Extending Claude/Cursor/other AI tools.
When not to
  • Simple apps that don't need tool use.
  • When provider-specific function calling is enough.
  • When security is paramount (MCP adds attack surface).

Definition

MCP is a protocol that lets AI models (Claude, GPT, etc.) discover and use tools/resources from external servers. A server exposes: tools (functions to call), resources (data to read), and prompts (templates). The client (AI) calls them. JSON-RPC over stdio or .

Why it exists

Every AI app builds its own integrations (tools, plugins). MCP standardizes this — write a server once, any AI can use it. Connect to databases, APIs, filesystems, GitHub, etc.

How it works

An MCP server exposes capabilities via a standard interface. The AI client discovers them, then calls tools (e.g., 'read_file', 'query_database') during a conversation. Servers run as subprocesses (stdio) or remote services (/SSE). Claude Desktop, , and others support it.

Scaling limits

New standard, still evolving. Security (running arbitrary tools). Performance (subprocess overhead). Not all AI providers support it yet. Server management complexity.

Cost shape

Free and open source. MCP servers are just code you run.

Alternatives

  • OpenAI function calling — Built-in, simpler, but OpenAI-only.
  • LangChain tools — Framework-specific, more complex.
  • Direct API calls — No standardization, more code.

Failure modes

  • Security (tool misuse).
  • Protocol evolving.
  • Subprocess management.
  • Not universal (provider support varies).

, , ,