Model Context Protocol (MCP)

MCP is an open-source standard for connecting AI applications to external systems.

Using MCP, AI applications like Claude or ChatGPT can connect to data sources (local files, databases), tools (search engines, calculators) and workflows (specialised prompts) — enabling them to access key information and perform tasks.

Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardised way to connect electronic devices, MCP provides a standardised way to connect AI applications to external systems.

Adapted for this playbook from the Model Context Protocol documentation (protocol revision path 2026-07-28). Official site: modelcontextprotocol.io. Selected diagrams from the MCP docs repository. This is not a substitute for the upstream docs — always verify against the current specification.

MCP as the standardised connector between AI hosts and external capabilities. Source: Model Context Protocol documentation / playbook diagram.

Source: modelcontextprotocol/docs hero-light.svg.

Executive summary

Use this section when you need a structured, enterprise-friendly walkthrough of MCP: what it is, how hosts/clients/servers relate, how to build and connect servers, how to debug with the Inspector, and how auth, extensions and the registry fit together — always with links back to the official MCP docs.

When to use this page

• Designing agent tooling that must work across Cursor, Claude, ChatGPT, VS Code and other MCP hosts.
• Exposing internal APIs, filesystems or SaaS systems as tools/resources without bespoke plugins per host.
• Briefing architects on protocol layers (data vs transport), local vs remote servers, and security posture.
• Pair with Agentic AI, Security Engineering and 🤗 Transformers.

What can MCP enable?

• Agents can access Google Calendar and Notion, acting as a more personalised AI assistant.
• Claude Code can generate an entire web app from a Figma design.
• Enterprise chatbots can connect to multiple databases across an organisation so users analyse data in chat.
• AI models can create 3D designs in Blender and send them to a 3D printer.

Why MCP matters

Audience	Benefit
Developers	Less custom integration work when building or connecting AI apps and agents
AI applications / agents	Access to an ecosystem of data sources, tools and apps
End users	More capable assistants that can use your data and take actions when needed

Broad ecosystem support

MCP is an open protocol supported across a wide range of clients and servers. AI assistants like Claude and ChatGPT, development tools like Visual Studio Code, Cursor, MCPJam, and many others support MCP — build once, integrate everywhere.

Source: Model Context Protocol docs (og-image.png).

How this playbook maps to official MCP docs

Playbook section	Official docs area	Start here
Learn · Architecture	Architecture overview	Architecture
Learn · Servers	Tools, resources, prompts	Server concepts
Learn · Clients	Elicitation, roots, sampling	Client concepts
Learn · Versioning	YYYY-MM-DD revisions	Versioning
Develop · Build a server	First server tutorial	Build a server
Develop · Build a client	Host / client tutorial	Build a client
Develop · Agent Skills	Guided server design	Agent Skills
Develop · Local servers	Claude Desktop + stdio	Connect local
Develop · Remote servers	Streamable HTTP	Connect remote
Develop · Client best practices	Scaling hosts	Best practices
SDKs	Official language SDKs	SDKs
Tools · Debugging	Debug toolkit	Debugging
Tools · Inspector	Interactive tester	Inspector
Security · Authorization	OAuth 2.1	Authorization
Security · Best practices	Attack surface	Security practices
Extensions	Optional capabilities	Extensions
Registry	Public server metadata	Registry
Specification	Normative protocol	Specification

Credit policy

Every page in this section:

• Credits the Model Context Protocol project and links to the official doc URL
• Hosts selected diagrams from the official docs repository (and playbook diagrams where noted)
• Rewrites content for playbook clarity (British spelling, enterprise notes) rather than dumping the full multi-language tutorial verbatim
• Does not replace upstream docs — verify against modelcontextprotocol.io

Learning path

1. Architecture → Server concepts → Client concepts
2. Build a server then connect it locally
3. Validate with the MCP Inspector
4. Harden with authorization and security best practices
5. Publish metadata via the Registry when ready

Official “Start building” links

• Build servers — expose your data and tools
• Build clients — applications that connect to MCP servers
• Build MCP Apps — interactive UIs inside AI clients
• Architecture — core concepts

Key takeaways

• MCP is the USB-C of AI integrations: one protocol, many hosts and servers.
• Hosts create one client per server; servers expose tools, resources and prompts.
• Prefer official SDKs and the Inspector before hand-rolling JSON-RPC.
• Security and consent are first-class — especially for remote servers and tool calls.

What to do next

Continue with Architecture overview, or jump to Build an MCP server.