MCP (Model Context Protocol)

MCP is an open protocol introduced by Anthropic in late 2024 for connecting AI agents to tools and data sources in a standardized way. Instead of every agent reinventing tool definitions, MCP defines a common JSON-based contract: a server exposes tools/resources/prompts, a client (the agent) discovers and calls them.

For local AI: MCP servers can run entirely locally, exposing your file system, terminal, GitHub repos, databases, etc. to a local agent like Claude Code or OpenCode. There's a growing registry of community MCP servers for Notion, Linear, Slack, Postgres, Cloudflare, and dozens more.

Why it matters: before MCP, every coding agent had its own tool framework. After MCP, a tool you build once works across Claude Code, Claude Desktop, Cursor (with MCP support), Continue, and any future MCP-aware client. It's the USB-C of AI tooling.

An operator wiring up a local coding agent wants it to query a Postgres database and read files from a project directory. Instead of writing custom tool-calling glue for each, they run two MCP servers locally — a Postgres MCP server and a filesystem MCP server — and point their agent client at both. The agent discovers the available tools (query_table, read_file, list_directory) at session start and calls them the same way regardless of which underlying model is driving the session. This matters for local-first setups: swap the backend LLM from Claude to a self-hosted Qwen 2.5 32B with function calling support, and the same MCP servers keep working unchanged, because the tool contract lives in the protocol layer, not in the model.