Model Context Protocol (MCP)

An open standard that gives AI models a universal interface to connect with external tools, data sources, and services.

Last updated: 2026-04-12

Overview

Without MCP, an LLM is a “brain in a jar” — capable of reasoning but unable to act on external systems. It cannot access files, query databases, check email, search the web, or interact with APIs unless the application hard-codes each integration.

MCP solves this by defining one universal protocol. Build a connection once (an MCP server) and any AI model that supports MCP can use it — without custom integration work per model or per app.

The key distinction:

  • Skills = teach the model HOW to do things (procedures, domain knowledge, judgment)
  • MCP = give the model ACCESS to external systems (data, tools, services)

Skills and MCP are complementary. A skill might describe how to do a DCF valuation (the procedure); an MCP server provides access to the financial data the skill needs (the raw material).

How It Works

MCP servers expose tools that the model can call. Each server wraps a specific system (a database, an API, a file system) and presents it as a set of callable functions. The model decides which tools to call based on the task; the MCP server executes and returns results.

Servers are installed separately from the model. Most follow the same setup pattern: install (npm/pip), configure in Claude settings with API keys, restart.

Ecosystem Categories

The MCP server ecosystem breaks into broad functional categories:

CategoryWhat it provides
Search & WebAI-optimized search, web crawling, URL fetching
File System & Local DataRead/write local files, SQLite/Postgres queries, Excel manipulation
Developer ToolsGitHub, Git, browser automation, Docker, Sentry, codebase memory
Productivity & CommunicationGoogle Drive, Gmail, Calendar, Slack, Notion, Linear
Data & AnalyticsSnowflake, BigQuery, Supabase, MongoDB
Infrastructure & DevOpsAWS, Cloudflare, Kubernetes, Vercel
AI & ModelsElevenLabs TTS, Hugging Face, Replicate
UtilityTime/timezone, persistent memory, task management

Key Points

  • Start small: install a role-appropriate starter pack of 4–5 servers; don’t install all at once
  • Official Anthropic servers are the simplest to set up — usually one command
  • fastmcp is the fastest way to build custom servers when nothing existing fits (Python, afternoon build time)
  • MCPHub manages multiple servers via a single dashboard — useful once you have 5+ running
  • Context7 deserves special mention: injects up-to-date library documentation into context, eliminating hallucinated APIs and deprecated method calls

Starter Packs by Role

RoleRecommended servers
DevelopersFilesystem + GitHub + Context7 + Codebase Memory + Sentry
Knowledge WorkersFilesystem + Google Drive + Gmail + Google Calendar + Notion
Data AnalystsFilesystem + SQLite + PostgreSQL + Excel + Tavily
Content CreatorsFilesystem + Tavily + Obsidian + markdownify + Slack
DevOpsFilesystem + Docker + GitHub + AWS + Kubernetes

Connections

  • claude-code-skills — skills and MCP are complementary: skills = HOW, MCP = ACCESS
  • thin-harness-fat-skills — MCP servers are part of the “fat” extension layer the harness delegates to
  • coding-agent — MCP servers are one of the tool categories available to a coding agent
  • mcp-servers — curated reference list of 40 notable servers

Sources