Hub · Updated
The MCP Guide: Model Context Protocol for Business
Learn · Build · Hire
The Model Context Protocol (MCP) is an open standard that lets AI assistants connect to your tools, databases and APIs through one universal interface. This guide covers the architecture, the server primitives, the growing ecosystem and what it means for your business.
Maintained by George Shvaya, founder of Webvello.
Where to start
Each track answers a different question. Pick the one that sounds like you.
Learn
MCP is new to you and you want to understand the protocol before you build anything.
Build
You have a use case and want to wire an MCP server to your data or your workflow.
Hire
You want MCP integration for your product or team and need a partner who has shipped it.
What MCP is
MCP stands for Model Context Protocol. Anthropic announced it on November 25, 2024 as “an open protocol that enables seamless integration between LLM applications and external data sources and tools.” It was created by David Soria Parra and Justin Spahr-Summers at Anthropic.
The current spec version is 2026-07-28.
The USB-C analogy
Before MCP, every AI application needed a custom connector for every data source — like having a different cable for every device. MCP is the USB-C of AI integration: one standard plug that lets any compliant host talk to any compliant server. Build your server once and it works in Claude Desktop, Claude Code, ChatGPT, VS Code, Cursor and any future host that speaks the protocol.
For the glossary definition, see What is MCP (Model Context Protocol)?
The architecture
MCP follows a client-server model with three distinct roles, communicating over JSON-RPC 2.0.
Hosts
AI applications the user interacts with directly. Claude Desktop, Claude Code, ChatGPT and VS Code are all hosts. A host creates and manages one or more MCP clients.
Clients
Connectors that live inside the host. Each client maintains a one-to-one connection with a single MCP server, handling protocol negotiation and message routing.
Servers
Programs that expose capabilities—tools, resources and prompts—to the AI through the protocol. A server can wrap a database, an API, a file system or any other data source.
Transports
MCP defines two transport mechanisms:
Stdio
Local, process-to-process communication. The host launches the server as a child process and communicates over standard input/output. Ideal for CLI tools and desktop applications.
Streamable HTTP
Remote transport using HTTP requests with optional Server-Sent Events (SSE) for streaming. Suited to cloud-hosted servers that multiple clients connect to over the network.
What servers can do
MCP servers expose capabilities through three core primitives. Additionally, the protocol defines one client primitive.
Tools
Server primitive
Executable functions the AI can invoke. A tool might query a database, call an external API, run a calculation or perform any action. The model decides when to call a tool based on the user’s request and the tool’s description.
Resources
Server primitive
Data sources the server makes available for the AI to read. Resources can be files, database records, API responses or any structured data. They are read by the application, not invoked by the model directly.
Prompts
Server primitive
Reusable prompt templates the server provides. They allow servers to offer pre-built workflows—for example a code review template or a data analysis prompt—that the user can select and the host fills in.
Elicitation
Client primitive
A mechanism that lets a server ask the user for input through the host. When a server needs a decision or a piece of information to continue, it sends an elicitation request and the host presents it to the user.
Extensions beyond the core
The MCP ecosystem is evolving with additional capabilities built on the protocol: Tasks enable async, long-running operations; MCP Apps allow servers to provide inline UI within the host; and Skills over MCP let servers expose higher-level reusable capabilities. Sampling was deprecated in the 2026-07-28 spec version.
Who supports MCP
The protocol launched with Anthropic’s own products and has since been adopted by competing vendors and developer tools.
| Application | Vendor | Role |
|---|---|---|
| Claude Desktop | Anthropic | Host |
| Claude Code | Anthropic | Host |
| ChatGPT | OpenAI | Host |
| VS Code (Copilot) | Microsoft | Host |
| Cursor | Cursor | Host |
| Zed | Zed Industries | Host (early adopter) |
| Replit | Replit | Host (early adopter) |
| Codeium / Windsurf | Codeium | Host (early adopter) |
| Sourcegraph Cody | Sourcegraph | Host (early adopter) |
Pre-built reference servers
Anthropic shipped reference servers at launch for common integrations:
- GitHub
- Git
- Google Drive
- Slack
- PostgreSQL
- Puppeteer
Early adopters
Companies that adopted MCP in its early days and contributed to the ecosystem:
- Block
- Apollo
- Zed
- Replit
- Codeium
- Sourcegraph
Want MCP working for your business?
Webvello builds custom MCP servers, wires them into Claude Code or your own AI agent pipeline, and handles the infrastructure so your team can focus on the product.
Building with MCP
The official specification, SDKs and community reference servers you need to start building.
- modelcontextprotocol.io — official specification and documentation
- github.com/modelcontextprotocol — SDKs and reference implementations
- Anthropic announcement — November 25, 2024
- How to build an MCP server — walkthrough
- What is MCP? Glossary definition
- AI Search Playbook — the broader context for AI integrations
MCP for your business
Practical ways the protocol turns AI assistants from chat windows into tools that act on your systems.
- AI agent development — build agents that use MCP to reach your data
- Claude Code development — custom toolchains with MCP servers
- Claude Code consulting — planning and architecture for MCP workflows
- AI crawler checker — see which agents can reach your site
- Schema generator — structured data AI systems can parse
- AEO audit — check your AI visibility baseline
Related Webvello services
Webvello builds MCP servers, AI agent pipelines and the surrounding infrastructure.
Security and trust
The MCP specification addresses security through principles rather than protocol-level enforcement. Implementations are responsible for applying these in practice.
User consent
The spec requires that users explicitly approve tool use. A host must not execute a tool or access a resource without the user’s knowledge and consent.
Data privacy
Servers should only expose data the user has authorized. Hosts should not send conversation context to servers without user awareness. Data minimization is encouraged.
Tool safety
Tools that perform actions (writing data, sending messages, making purchases) must be treated with extra caution. The spec recommends clear descriptions and confirmation flows.
No protocol-level enforcement
MCP itself does not mandate authentication, authorization or encryption. Each implementation decides how to secure the transport, validate identity and control access.
Common questions
What is the Model Context Protocol?
MCP is an open protocol, announced by Anthropic on November 25, 2024, that provides a standardized way for AI applications to connect to external data sources and tools. It uses JSON-RPC 2.0 and defines three server primitives—tools, resources and prompts—so any compliant host can talk to any compliant server without custom integration code.
Who created MCP?
MCP was created by David Soria Parra and Justin Spahr-Summers at Anthropic. The specification is open source and hosted at modelcontextprotocol.io, with SDKs and reference servers on GitHub under the modelcontextprotocol organization.
How is MCP different from a regular API?
A regular API exposes one service with its own authentication, schema and calling convention. MCP is a protocol layer on top of APIs: it gives every server the same interface (tools, resources, prompts over JSON-RPC 2.0) so a host like Claude Desktop or VS Code can connect to any server without knowing its internals. Think of it as USB-C for AI integrations—one plug, many devices.
Which AI assistants support MCP?
As of the current spec (2026-07-28), MCP is supported by Claude Desktop, Claude Code, ChatGPT, VS Code with Copilot and Cursor. Early adopters building on MCP include Zed, Replit, Codeium and Sourcegraph. The protocol is open, so any application can add support.
What are the two MCP transport methods?
MCP defines two transports. Stdio is local, process-to-process communication suited to CLI tools and desktop apps. Streamable HTTP is a remote transport that uses HTTP requests with optional Server-Sent Events (SSE) for streaming, suited to cloud-hosted servers.
Is MCP secure?
The specification requires user consent before any tool executes, mandates data privacy considerations and emphasizes tool safety. However, MCP does not enforce security at the protocol level—each implementation is responsible for authentication, authorization and access control on top of the protocol.
Can Webvello build an MCP server for my business?
Yes. Webvello builds custom MCP servers, connects them to your databases, APIs and internal tools, and deploys them inside Claude Code, Claude Desktop or your own AI agent pipeline. Start with a free growth plan to scope the integration.
What happened to Sampling in MCP?
Sampling was deprecated in the 2026-07-28 spec version. The pattern it served—servers requesting the host to run a model completion—is being reconsidered. The current spec instead introduces Elicitation as a client primitive, which lets a server ask the user for input through the host.
Ready to build with MCP?
Webvello builds custom MCP servers, connects them to your data and deploys them inside your AI workflows. Start with a free growth plan to scope the integration.