MCP
Model Context Protocol
Scroll to explore
ChatGPTClaudeGeminiCursorStripeHubSpotSlackGitHubPostgresN × M = 20 custom integrations
ChatGPTClaudeGeminiCursorMCPStripeHubSpotSlackGitHubPostgresN + M = 9 standard connections
mcp-protocol
1. INITIALIZE
Client ↔ Server exchange capabilities
2. DISCOVER
tools/list → [{name: "hubspot_search", ...}]
3. EXECUTE
tools/call("hubspot_search", {query: "deals"})
4. RESPOND
{deals: [{name: "Acme", amount: "$50K"}]}
AI Assistant
What were our top deals this quarter?
hubspot_search_deals
stripe_list_payments

Top 5 deals this quarter:

Acme Corp$52K • Paid
Globex Inc$38K • Paid
Initech LLC$27K • Pending
AI + MCP App
Show me a dashboard of this quarter
Q1 2026 Dashboard
$127K
Revenue
23
Deals
↑ 18%
Growth
97M+
Monthly SDK Downloads
10K+
Public MCP Servers
ChatGPTClaudeGeminiCursorVS CodeCopilotWindsurf
Governed by the Linux Foundation
The Protocol

What is MCP?

The Model Context Protocol is the USB-C for AI — one open standard that connects any AI assistant to any tool, database, or API. Created by Anthropic, now governed by the Linux Foundation.

The Problem

N × M Was Unsustainable

Before MCP, 10 AI models needing 20 tools meant building 200 separate integrations. Each one fragile, vendor-locked, and expensive to maintain. Every platform had its own plugin system — nothing was portable.

The Standard

One Protocol. Every Platform.

MCP transforms N×M into N+M. Build a tool server once — it works in ChatGPT, Claude, Cursor, VS Code, Windsurf, and every future MCP-compatible client.

How It Works

Connect. Discover. Execute.

MCP uses JSON-RPC 2.0 between AI clients and tool servers. On connection, the client discovers every available tool with its input schema. The AI decides what to call and when — no pre-built workflows needed.

In Action

Follow a Real Request

“What were our top deals?” The AI discovers HubSpot and Stripe tools, calls them automatically, cross-references deal data with payments, and delivers a formatted answer. One question — MCP handles the rest.

MCP Apps

Interactive UI Inside AI

Launched January 2026, MCP Apps let tools return rich interactive interfaces — dashboards, forms, charts — rendered directly inside the AI conversation. Your chat becomes a workspace.

The Ecosystem

From Zero to Universal

97 million monthly SDK downloads. 10,000+ public servers. Adopted by every major AI platform. Governed by the Linux Foundation’s Agentic AI Foundation, co-founded by Anthropic, OpenAI, and Block.

The Specification

Three Core Primitives

MCP servers expose three types of capabilities. Together they give an AI everything it needs to understand and act on your data.

Tools

Functions the AI can call to take action — "create a Stripe invoice," "search HubSpot contacts," "run a SQL query." Each tool has a JSON Schema defining its input parameters, so the AI knows exactly what arguments to provide.

Tools include annotations like readOnlyHint and destructiveHint so clients can display appropriate confirmation prompts.

Resources

Read-only data the AI can pull into its context — files, database schemas, API documentation, configuration. Resources are identified by URIs (like file:///schema.sql) and can be text or binary.

Servers can notify clients when resources change, allowing the AI to stay up-to-date with live data automatically.

Prompts

Reusable templates that combine instructions with tool-calling workflows. A server can expose a "quarterly review" prompt that teaches the AI the best sequence of calls to pull and format your data.

Prompts can accept arguments and include multi-turn message sequences, giving servers fine-grained control over complex workflows.

Transport & Security

Streamable HTTP

MCP servers communicate over HTTP with server-sent events (SSE) for streaming responses. Servers can be hosted anywhere — no local processes or Docker containers required. MCPBundles hosts all servers for you at mcp.mcpbundles.com.

OAuth 2.1 Authentication

The MCP spec includes OAuth 2.1 for server authentication, with PKCE and dynamic client registration. MCPBundles manages OAuth flows and API keys for every provider — connect your credentials once, and every tool call is authenticated automatically.

New in 2026

MCP Apps — Interactive UI Inside AI

MCP Apps is the first official extension to MCP, launched January 2026. Instead of returning plain text, tools can now return rich interactive interfaces — dashboards, forms, charts, and workflows — that render directly inside the AI conversation.

How MCP Apps Work

Tool Returns UI

A tool includes a ui:// resource URI in its response metadata, pointing to bundled HTML and JavaScript.

Sandboxed Render

The host (ChatGPT, Claude, etc.) fetches the resource and renders it in a secure sandboxed iframe — no access to parent page or cookies.

Bidirectional Data

The app and AI communicate via JSON-RPC over postMessage. The UI can call tools, and the host can push fresh results to the app.

The result: your AI conversation becomes an interactive workspace, not just a text chat.

Dashboards

Live charts and metrics from Stripe, PostHog, or your database — rendered inline with drill-down capability.

Forms & Wizards

Multi-step configuration flows, data entry forms, and approval workflows without leaving the conversation.

Data Exploration

Filter, sort, and drill into datasets interactively rather than asking the AI to re-query on every change.

Real-Time Monitoring

Live metrics, error rates, and deployment statuses that update in real time within the chat interface.

Supported MCP Apps Clients

ChatGPTClaudeGooseVS CodePostman+ more coming
Comparison

How MCP Compares

MCP is the only approach that is both AI-native and standardized

ApproachSetupAI-NativeAuthUI
MCPOne URL per serverBuilt for AI agentsOAuth 2.1 / API keysMCP Apps
REST APIsPer-endpoint codeHuman-orientedVaries per APINone
Function CallingPer-provider codeAI-aware, not standardCustom per integrationNone

MCP is the only approach that combines AI-native tool discovery (the agent decides what to call), a standardized protocol (one integration works everywhere), and interactive UI (MCP Apps render dashboards inside the conversation).

Ecosystem

Adoption & Governance

MCP went from a new protocol to the industry standard in 14 months. It's now governed by the Linux Foundation with backing from every major AI company.

November 2024

Anthropic Open-Sources MCP

Anthropic releases the Model Context Protocol as an open-source standard, with SDKs for Python and TypeScript and initial support in Claude Desktop.

2025

Industry-Wide Adoption

OpenAI adds MCP to ChatGPT. Google integrates MCP into Gemini. Microsoft adds support in Copilot and VS Code. Thousands of MCP servers appear across the ecosystem.

December 2025

Agentic AI Foundation

Anthropic donates MCP to the Linux Foundation's new Agentic AI Foundation (AAIF), co-founded by Anthropic, OpenAI, and Block with Platinum members AWS, Google, and Microsoft.

97M+

Monthly SDK downloads

10K+

Public MCP servers

6+

Major AI platforms

7

AAIF founding members

Compatibility

Supported AI Platforms

MCP works across all major AI assistants and developer tools

AI

ChatGPT

OpenAI

C

Claude

Anthropic

G

Gemini

Google

Cu

Cursor

AI code editor

VS

VS Code

Microsoft

W

Windsurf

Codeium

Co

Copilot

Microsoft

+

More

Growing ecosystem

Why It Matters

Why MCP Matters for You

One Standard, Every Platform

Build or connect to an MCP server once and it works in ChatGPT, Claude, Cursor, VS Code, Windsurf, Gemini, and every future MCP-compatible client. No vendor lock-in.

The AI Decides

Unlike workflow builders where you define every step, MCP gives the AI a toolbox and lets it figure out the right sequence. Ask for a quarterly report and the AI pulls from Stripe, HubSpot, and PostHog automatically.

Secure by Design

OAuth 2.1 with PKCE, encrypted credentials at rest, workspace-level access control, tool annotations for destructive actions, and full execution audit logs. Your data stays under your control.

What MCPBundles Adds

MCP is the protocol. MCPBundles is the platform that makes it effortless. We host hundreds of pre-built MCP servers, manage OAuth credentials, combine tools from multiple services into single endpoints, and provide workspace-level access control — so you get the power of MCP without running anything yourself.