Skip to main content

​
What is MCP (Model Context Protocol)?

MCP (Model Context Protocol) is an open standard that defines how AI assistants connect to external tools, data sources, and services. It was created by Anthropic in November 2024 and donated to the Linux Foundation in March 2025, making it a vendor-neutral protocol that any AI client or service provider can implement. Before MCP, connecting an AI assistant to an external service required custom integration code for each combination of AI client and service. If you wanted Claude to access your CRM, you built a custom integration. If you wanted ChatGPT to access the same CRM, you built a different custom integration. For every new AI client, you repeated the work. MCP eliminates this fragmentation. A service builds one MCP server, and it works with every MCP-compatible AI client — Claude, ChatGPT, Cursor, VS Code, Windsurf, and dozens more. An AI client implements MCP support once, and it works with every MCP server. The protocol standardizes how AI models discover available tools, call them with structured inputs, and receive structured outputs. Think of MCP as USB-C for AI. Before USB-C, every device had its own charging cable. USB-C standardized the connector so one cable works with everything. MCP does the same for AI-to-tool connections.

​
How MCP Works

MCP defines a communication protocol between two parties: clients (AI assistants) and servers (tools and services). The client discovers what the server can do, then calls those capabilities as needed during conversations.

​
The Three Primitives

MCP exposes three types of capabilities from servers to clients:
Tools are functions that the AI can call to take actions or retrieve information. Each tool has a name, a description (so the AI knows when to use it), and a defined input schema (so the AI knows what parameters to provide).Examples:
  • list_contacts — Search and filter contacts in a CRM
  • create_appointment — Book an appointment on a calendar
  • get_weather — Retrieve current weather for a location
  • run_query — Execute a database query
When a user asks the AI to do something that requires an external tool, the AI selects the appropriate tool, fills in the parameters, and sends a structured request to the MCP server. The server executes the action and returns the result.
Resources are read-only data that the AI can access for context without making explicit function calls. They are similar to files or documents that the AI can reference during conversations.Examples:
  • smartalex://platform-status — Current platform health and version
  • github://repo/readme — The README file of a repository
  • database://schema — The schema of a database
Resources help the AI understand the current state of a system without needing to call a tool. They are typically used for context-setting at the beginning of a conversation.
Prompts are pre-built conversation templates that guide the AI through complex multi-step tasks. They are like expert playbooks that the AI follows when the user selects them.Examples:
  • campaign-strategy — Walk through designing an outbound calling campaign
  • code-review — Analyze code changes for bugs, style issues, and security concerns
  • data-analysis — Guide a structured analysis of a dataset
Prompts combine instructions, tool usage, and conversation flow into reusable templates. They ensure consistent, high-quality outputs for common workflows.

​
Communication Flow

1. Client connects to Server
2. Server advertises its capabilities (tools, resources, prompts)
3. User asks the AI to do something
4. AI selects the appropriate tool and constructs a request
5. Client sends the request to the Server
6. Server executes the action and returns the result
7. AI incorporates the result into its response to the user

​
Transport Methods

MCP supports two transport methods:
TransportHow It WorksBest For
stdioServer runs locally as a process, communicates via standard input/outputDesktop AI clients (Claude Desktop, Cursor, VS Code)
HTTP (Streamable HTTP)Server runs remotely, communicates via HTTPSCloud AI clients (ChatGPT, web-based assistants), hosted services
The stdio transport is simpler and requires no network configuration — the AI client launches the server process directly. The HTTP transport enables cloud-hosted MCP servers that any client can connect to over the internet.

​
Why MCP Matters

​
The Integration Problem Before MCP

Before MCP, connecting AI to external tools required:
  1. Custom code per integration — Every AI client needed its own connector for every service
  2. Provider lock-in — Integrations built for one AI client did not work with others
  3. Maintenance burden — Each connector needed separate updates, authentication handling, and error management
  4. Discovery problem — No standard way for AI to know what tools are available
For a platform with 20 tools and 5 AI clients, this meant building and maintaining 100 separate integrations.

​
What MCP Changes

With MCP:
  1. Build once, work everywhere — One MCP server works with Claude, ChatGPT, Cursor, VS Code, and any future MCP client
  2. No lock-in — Switch AI clients without rebuilding integrations
  3. Automatic discovery — AI clients discover available tools, their descriptions, and their input schemas automatically
  4. Standard authentication — Consistent auth patterns across all integrations
  5. Growing ecosystem — Thousands of MCP servers are already available, with more published daily
For a platform with 20 tools and 5 AI clients, this means building 1 MCP server instead of 100 integrations.

​
Industry Adoption

MCP has been adopted rapidly across the AI industry:
  • Anthropic — Claude Desktop, Claude Code (creator of MCP)
  • OpenAI — ChatGPT supports MCP for tool calling
  • Google — Gemini integration announced
  • Microsoft — VS Code with GitHub Copilot, Copilot Studio
  • Cursor — Full MCP support for code editing
  • Windsurf — MCP support for development workflows
  • Block (Square) — Enterprise MCP adoption
  • Apollo, Replit, Sourcegraph, Zed — And many more
The Linux Foundation governance ensures MCP evolves as an open standard rather than being controlled by any single company.

​
MCP vs REST APIs

MCP and REST APIs solve related but different problems. REST APIs are designed for application-to-application communication. MCP is designed for AI-to-application communication.
DimensionMCPREST API
Designed forAI assistants calling toolsApplications calling services
DiscoveryAutomatic — AI reads tool descriptions and schemasManual — developer reads documentation
Input constructionAI fills parameters from natural languageDeveloper writes code to construct requests
Error handlingAI interprets errors and adaptsDeveloper writes error handling code
AuthenticationStandard patterns (API key, OAuth) built into protocolVaries per API
Multi-step workflowsAI chains tool calls based on conversation contextDeveloper writes orchestration code
Schema evolutionServer advertises current capabilities dynamicallyVersioned endpoints, breaking change management
Who uses itAI models (via clients)Developers (via code)
MCP does not replace REST APIs. Most MCP servers are built on top of existing REST APIs. The MCP server acts as a translation layer that makes a REST API accessible to AI assistants in a standardized way. 
User → AI Client → MCP Server → REST API → Service
If you already have a REST API, building an MCP server on top of it typically takes hours, not weeks. The MCP server wraps your existing API endpoints as tools, adding descriptions and schemas that help AI clients understand when and how to use them.

​
Supported Clients

Any AI client that implements the MCP protocol can connect to any MCP server. The ecosystem is growing rapidly.

​
Desktop AI Clients

ClientTransportNotes
Claude DesktopstdioFirst-party support from Anthropic
Claude Codestdio, HTTPCLI-based, supports both local and cloud MCP
CursorstdioPopular AI code editor
VS Code (GitHub Copilot)stdioVia MCP extension or built-in support
WindsurfstdioAI-powered development environment
ZedstdioFast, collaborative code editor

​
Cloud AI Clients

ClientTransportNotes
ChatGPTHTTPOpenAI’s consumer AI assistant
Claude.aiHTTPAnthropic’s web-based assistant
Google GeminiHTTPGoogle’s AI assistant

​
Development Tools

ToolTransportNotes
Copilot StudioHTTPMicrosoft’s enterprise AI builder
ReplitstdioCloud development environment
Sourcegraph CodystdioAI code assistant

​
Building an MCP Server

Building an MCP server involves defining the tools, resources, and prompts your service exposes, then implementing the handlers that execute when an AI client calls them.

​
Minimal Example (TypeScript)

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";

const server = new McpServer({
  name: "my-service",
  version: "1.0.0",
});

// Define a tool
server.tool(
  "get_customer",
  "Look up a customer by email address",
  { email: z.string().email() },
  async ({ email }) => {
    const customer = await db.customers.findByEmail(email);
    return {
      content: [{ type: "text", text: JSON.stringify(customer) }],
    };
  }
);

// Start the server
const transport = new StdioServerTransport();
await server.connect(transport);

​
Official SDKs

MCP provides official SDKs in multiple languages:
LanguagePackageStatus
TypeScript/JavaScript@modelcontextprotocol/sdkStable
PythonmcpStable
Java/Kotlinio.modelcontextprotocol:sdkStable
C#/.NETModelContextProtocolStable
Gogithub.com/mark3labs/mcp-goCommunity
RustrmcpCommunity
Swiftmcp-swift-sdkCommunity

​
Key Design Principles

When building an MCP server:
  1. Write clear tool descriptions — The AI uses these to decide when to call your tool. Vague descriptions lead to misuse.
  2. Use specific input schemas — Define required and optional parameters with types and constraints. The AI constructs better inputs from well-defined schemas.
  3. Return structured data — JSON responses are easier for the AI to interpret than free-form text.
  4. Handle errors gracefully — Return meaningful error messages that help the AI recover or inform the user.
  5. Keep tools focused — One tool per action. A list_contacts tool and a create_contact tool are better than a single manage_contacts tool.

​
SmartAlex’s MCP Server

SmartAlex provides a fully-featured MCP server that lets AI assistants manage the entire voice agent platform through natural conversation. The SmartAlex MCP server includes:
  • 18 tools across contacts, agents, campaigns, calls, deals, webhooks, and platform management
  • 7 resources for read-only context (platform status, active campaigns, recent calls, deal pipeline)
  • 5 prompts for guided workflows (campaign strategy, lead qualification, agent design, call analysis, pipeline review)
  • 7 pre-built workflows that chain tools for common operations

​
What You Can Do

With SmartAlex’s MCP server connected to your AI assistant, you can manage your voice agent platform through conversation:
  • “Show me all contacts tagged as ‘hot lead’ who called this week”
  • “Create a new voice agent for appointment scheduling with a professional female voice”
  • “What is the conversion rate for the Q1 outbound campaign?”
  • “Set up a webhook to notify Slack when a call is completed”
  • “Create a deal for Acme Corp at $50,000 in the proposal stage”

​
Connection Options

MethodCommandBest For
Local (stdio)npx @smartalex/mcp-serverClaude Desktop, Cursor, VS Code
Cloud (HTTP)Connect to mcp.getsmartalex.comChatGPT, cloud-based clients
For detailed setup instructions, see the SmartAlex MCP Getting Started guide.

​
The MCP Ecosystem

The MCP ecosystem has grown rapidly since the protocol’s release. Thousands of MCP servers are available for virtually every category of software.

​
Common MCP Server Categories

CategoryExamples
Developer toolsGitHub, GitLab, Linear, Jira, Sentry
DatabasesPostgreSQL, MySQL, MongoDB, Supabase, Neon
CommunicationSlack, Gmail, Outlook, Discord
CRM & SalesSalesforce, HubSpot, Pipedrive
Cloud infrastructureAWS, Google Cloud, Cloudflare, Vercel
AI & VoiceSmartAlex, ElevenLabs, OpenAI
File storageGoogle Drive, Dropbox, S3, R2
KnowledgeNotion, Confluence, Wikipedia, web search
FinanceStripe, QuickBooks, Plaid
DesignFigma, Canva

​
Finding MCP Servers

  • MCP Server Registry — Official list maintained by the MCP project
  • Smithery — Community directory of MCP servers
  • Glama — Searchable MCP server catalog
  • npm / PyPI — Search for mcp-server-* packages

​
Frequently Asked Questions

No. MCP was created by Anthropic but donated to the Linux Foundation as an open standard. It is supported by ChatGPT (OpenAI), Gemini (Google), Cursor, VS Code, Windsurf, and many other AI clients. Any AI assistant that implements the MCP protocol can connect to any MCP server. The protocol is vendor-neutral by design.
OpenAI function calling is a proprietary feature that lets GPT models call functions you define within the OpenAI API. It is specific to OpenAI models and requires OpenAI’s API format. MCP is an open protocol that works across all AI clients and models. MCP also adds resource and prompt primitives, automatic tool discovery, and standardized transport. Think of function calling as a feature of one AI provider. MCP is a protocol that works with all of them.
To build an MCP server, yes — you need programming skills. To use an MCP server, no. End users simply connect pre-built MCP servers to their AI assistant (often by pasting a configuration snippet or clicking an install button) and then interact with the tools through natural language conversation. The AI handles all the technical details of calling tools and interpreting results.
MCP supports standard authentication mechanisms (API keys, OAuth 2.0) and communicates over encrypted channels (HTTPS for HTTP transport). The security of an MCP connection depends on the server implementation and the authentication method used. Best practices include using scoped API keys, rotating credentials, and granting the minimum necessary permissions. MCP itself does not introduce security vulnerabilities beyond what the underlying API already has.
No. MCP servers only run when explicitly connected and authenticated by the user. The AI client must be configured with the MCP server’s address and credentials. Each tool call is initiated by the AI on behalf of the user, and the server enforces its own authorization rules. Users can disconnect MCP servers at any time. Additionally, most AI clients show tool calls to the user before executing them, providing transparency and control.
LangChain is a framework for building LLM applications in Python and JavaScript. LangChain tools are functions defined within a LangChain application. MCP is a protocol for connecting any AI client to any service. LangChain tools are application-specific. MCP servers are universal. However, LangChain has added MCP support, meaning LangChain applications can now use MCP servers as tool sources, combining LangChain’s orchestration with MCP’s universal connectivity.
Build an MCP server that wraps your existing API. If you have a REST API, each API endpoint becomes an MCP tool with a description and input schema. The official TypeScript and Python SDKs make this straightforward — a basic MCP server with 5-10 tools can be built in a few hours. Publish it as an npm or PyPI package so users can install it with one command.
MCP is governed by the Linux Foundation with contributions from Anthropic, OpenAI, Google, Microsoft, and the open-source community. Active development areas include improved authentication (OAuth 2.1), richer resource types, better error handling, streaming responses, and server-to-server communication. As more AI clients and services adopt MCP, it is becoming the de facto standard for AI-to-tool integration, similar to how REST became the standard for web APIs.

​
Learn More

MCP Specification

Read the full MCP specification, including protocol details, transport methods, and capability negotiation.

SmartAlex MCP Server

Connect your AI assistant to SmartAlex’s 18-tool MCP server for managing voice agents, campaigns, contacts, and analytics through natural conversation.