A standard protocol for LLM apps to talk to tools, resources, and data.

MCP — Model Context Protocol

MCP is to LLM agents what LSP is to IDEs: one client (Claude Desktop, Copilot, Cursor) speaks to many servers (filesystem, GitHub, Postgres, custom). Built on JSON-RPC over stdio or SSE, it standardizes tools, resources, prompts, and sampling so any model can integrate any capability.

Interactive Pipeline Flow

IngestProcessServe

📐 Define — tools + resources

Define your server's tools (callable functions), resources (readable data), and prompts (templates) with JSON Schema.

Architecture Diagrams

Server lifecycle: build, sign, distribute, version.

Loading diagram...

Key Design Decisions

🔌Build for Stdio First

Stdio transport works everywhere (no network, no auth). Add SSE/HTTP only when remote hosting is required.

📐Strict Schemas

Inputs and outputs need JSON Schema. The LLM uses these to format calls — vague schemas → wrong arguments.

🛡️Treat as Privileged

MCP servers run with user permissions. Validate all args; never trust LLM-generated tool input blindly.

📦Distribute via Package

Publish to npm/PyPI; users add via uvx or npx. Keeps install one-line; you control updates.

🧪mcp-inspector

The official inspector tool catches schema errors and round-trip issues before clients see them.

📡Stream Long Ops

For multi-second tool calls, send progress notifications. UX wins — user sees activity, not a frozen spinner.

Discussion