Model Context Protocol: the open standard connecting AI to the world

Executive overview

AI models are isolated by default — they can't access live data, external tools, or services they weren't trained on. Model Context Protocol (MCP) is an open standard created by Anthropic that gives any AI model a universal way to connect to external data sources and tools.

Think of it as USB-C for AI: one connector standard instead of dozens of custom integrations. Rather than every developer rebuilding bespoke connections, they build once to the MCP standard and share it with everyone.

The core insight: standardisation unlocks a flywheel — more MCP servers attract more clients, which attracts more servers, accelerating the path to autonomous agents.

The problem MCP solves

• AI models are "caged" — siloed from live data and real-world tools
• Developers kept building custom integrations in isolation, reinventing the wheel repeatedly
• No shared learnings, wasted resources, slow collective progress
• OpenAI's earlier plugin approach (ChatGPT plugins, function calling) was closed — only usable with OpenAI models

Three phases of AI tool connectivity

1. Phase 1 — Caged AI: Models trained on static data, no external reach
2. Phase 2 — Custom integrations: Developers build bespoke connections in silos; still the dominant reality today
3. Phase 3 — Open standard (MCP): Build once, share with all; the transition is underway

How MCP works

• Client: An application with an AI model embedded (e.g. Claude Desktop, Cursor IDE)
• MCP server: A translator sitting between the AI client and an external tool or dataset
• The client speaks the standardised MCP protocol; the server translates it into the API language of the target service
• Servers can reach local machine resources (files, databases) or remote services (Slack, email, calendars, search)
• Human-in-the-loop approval gates are supported before sensitive actions execute

Why Anthropic open-sourced it

• Classic "second place" strategy in tech: create an open standard to accelerate adoption and close the gap with the leader
• Precedents: Google open-sourced Kubernetes (to challenge AWS/Azure), Android (to challenge iPhone), Meta released Llama (to challenge OpenAI/Anthropic)
• Anthropic is perceived as behind OpenAI; an open, vendor-neutral standard lets the entire ecosystem build on their infrastructure
• OpenAI's closed plugin approach didn't achieve broad adoption; MCP's openness is the differentiator

MCP ecosystem today

• Anthropic maintains a reference list of official servers: Google Drive, GitHub, Brave Search, Slack, Google Maps, and more
• Community marketplaces (mcp.so, glama.ai, Kline) host hundreds of community-built servers
• Early-stage: setup still requires manual config file edits and API key management — not yet consumer-friendly
• Servers can be finicky; the standard is still maturing

Roadmap: first half of 2025

• Security: Replace API token/password auth with OAuth 2.0 — the standard used across most modern web apps
• Discovery: AI auto-detects which MCP servers are relevant for a task, configures them automatically, then requests user approval
• Remote servers: Move MCP servers from local machines to the cloud, reducing setup friction and increasing throughput for agent workloads