LIVE
Which tools do Claude, Codex and Cursor choose? We measured 17k runs to find out03/09/26 · Anthropic|OpenAI's GPT-6 Astra on ARC-AGI-303/09/26 · OpenAI|GPT-6 Astra03/09/26 · OpenAI|OpenAI begins rolling out GPT-6 Astra03/09/26 · OpenAI|ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize03/09/26 · Anthropic|Sparks Fly: NVIDIA Accelerates Local AI at IFA 202603/09/26 · NVIDIA|Introducing WeatherNext 3, our most advanced and accurate global weather AI model03/09/26 · Google|Porting my 1993 Amiga game to Godot, with an LLM reading the 68000 assembly03/09/26 · Anthropic|Claude outage – Resolved03/09/26 · Anthropic|Daybreak for Frontline Defenders: $1B to protect essential services03/09/26 · OpenAI|NeoMME: an efficient Multimodal-native and Multilingual Encoder03/09/26 · Hugging Face|‘NBA 2K27’ With NVIDIA DLSS 5 Leads 28 New Games Coming to GeForce NOW03/09/26 · NVIDIA|Which tools do Claude, Codex and Cursor choose? We measured 17k runs to find out03/09/26 · Anthropic|OpenAI's GPT-6 Astra on ARC-AGI-303/09/26 · OpenAI|GPT-6 Astra03/09/26 · OpenAI|OpenAI begins rolling out GPT-6 Astra03/09/26 · OpenAI|ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize03/09/26 · Anthropic|Sparks Fly: NVIDIA Accelerates Local AI at IFA 202603/09/26 · NVIDIA|Introducing WeatherNext 3, our most advanced and accurate global weather AI model03/09/26 · Google|Porting my 1993 Amiga game to Godot, with an LLM reading the 68000 assembly03/09/26 · Anthropic|Claude outage – Resolved03/09/26 · Anthropic|Daybreak for Frontline Defenders: $1B to protect essential services03/09/26 · OpenAI|NeoMME: an efficient Multimodal-native and Multilingual Encoder03/09/26 · Hugging Face|‘NBA 2K27’ With NVIDIA DLSS 5 Leads 28 New Games Coming to GeForce NOW03/09/26 · NVIDIA|
Intermediate🔌

MCP (Model Context Protocol): the standard that finally connects AI to your tools

The Model Context Protocol is the USB-C of AI. One standard to connect Claude, ChatGPT, Gemini to your 50 tools (Slack, GitHub, Notion, Drive...) without coding 50 integrations. Adopted by Anthropic, OpenAI, Google in 2026. Why it's the silent revolution.

14 min readPublished May 12, 2026· Updated August 27, 2026

In one sentence

MCP (Model Context Protocol) is an open standard that allows AIs (Claude, ChatGPT, Gemini...) to talk to your tools (Slack, Notion, GitHub, Drive, databases...) without coding an integration per tool. It became the de facto standard in 2026.

🔌
The analogy that works
Before MCP: each AI had its own method to connect to tools. Connecting Claude to Slack = 1 code. To GitHub = another code. To Notion = yet another. Multiply by 5 AIs × 50 tools = 250 integrations to maintain. Catastrophe. With MCP: one universal protocol. You write 1 MCP server for Slack, and all AIs can use it. Like USB-C that charges all your devices, regardless of brand.
The trap to avoid in 2026
Many companies still think MCP is "just for devs". Wrong. It's become a business topic: not exposing your SaaS via MCP in 2026 is like not having a REST API in 2015. Your customers will ask for it, and go elsewhere.

The problem MCP solves

Before MCP, the enterprise AI world was a chaos of integrations. Each AI vendor developed its own connectors: ChatGPT had its "plugins", Claude its "tools", Google its "extensions". No compatibility.

Concretely: you wanted to connect your favourite AI to your internal Slack? You coded a custom integration. You wanted to change AI 6 months later? You recoded everything. You wanted 3 AIs in parallel? 3× the work.

It's exactly the problem smartphones had with chargers before USB-C: an Apple cable, a Samsung cable, a mini-USB, a micro-USB... A mess.

Before MCP vs With MCP: the revolution
❌ BEFORE MCP N × M custom integrations Claude ChatGPT Gemini Slack GitHub Notion 9 integrations to code, maintain, debug ✅ WITH MCP 1 universal protocol Claude ChatGPT Gemini MCP protocol Slack GitHub Notion 6 connections to ONE MCP hub With 10 AIs × 50 tools: 500 integrations 60 connections to maintain manually thanks to a single standard → Time saved: 88%
MCP transforms the chaos of integrations into a universal standard.

MCP architecture: 3 components

The architecture is deliberately simple, which explains its rapid adoption. Three actors communicate:

🏗️
The 3 essential building blocks
1. Host = the application using the AI (Claude Desktop, ChatGPT, Cursor, your own app) 2. Client = the bridge between the Host and the Server (often managed automatically by the Host) 3. Server = the exposed tool (Slack, GitHub, your database, your CRM...) The Host asks a question to the Client, which sends it to the Server. The Server responds. The AI receives the data and formulates its response to the user. That's it.
📚Going deeper

How the dialogue works concretely

Imagine you ask Claude: "How many open tickets do I have in Jira?"

  1. Claude (Host) understands it needs to query Jira
  2. Claude asks the MCP Client to call the Jira server with list_tickets(status=open)
  3. The Jira MCP Server receives the request, calls the Jira API, retrieves "23 tickets"
  4. Server sends back "23 open tickets" to the Client
  5. Claude formulates the final response: "You have 23 open tickets in Jira, including 5 critical ones..."

All this in under 1 second, without you writing a single line of Jira code.

Supported transports

MCP supports several communication channels:

  • stdio: for local servers (fast, secure)
  • HTTP/SSE: for remote servers (cloud, multi-user)
  • WebSocket: for bidirectional real-time (rare)

99% of use cases = stdio (local) or HTTP (cloud). The SDKs handle transport automatically.

The lightning adoption: the timeline

MCP came out in November 2024. In 18 months, all major players adopted it. It's an unprecedented pace in the tech industry.

MCP adoption since its creation

  1. 🚀 Anthropic launches MCP

    Creation of the open source standard. First SDKs (Python, TypeScript). 10 reference servers (filesystem, postgres, slack...).

  2. 📦 Marketplace launched

    modelcontextprotocol.io publishes an official registry. 50 public servers available in 1 month.

  3. 🛠️ SaaS adoption

    Block, Shopify, Sourcegraph, Replit, Apollo publish their official MCP servers. The community contributes.

  4. 💼 Enterprise wave

    Fortune 500 companies start exposing their internal tools via MCP. 100+ public servers.

  5. 🎯 OpenAI joins

    ChatGPT + Agents support MCP natively. Thunderbolt: Anthropic and OpenAI aligned on the same standard.

  6. 🌐 Google aligns

    Gemini Enterprise announces MCP support. Microsoft Copilot Studio follows suit.

  7. 🏆 De facto standard

    250+ public servers. All major AI vendors support MCP. It's become a B2B prerequisite.

Who's already using it? (May 2026)

Adoption has been lightning-fast. Here's the state of the market in May 2026:

MCP adoption by AI vendor (May 2026)

Anthropic (creator)100%
OpenAI95%
Google (Gemini)85%
Microsoft80%
Mistral AI60%
xAI (Grok)40%
📚Going deeper

Popular MCP servers (marketplace top 10):

  1. GitHub — manage issues, PRs, code
  2. Slack — read/send messages
  3. Notion — search your docs
  4. Google Drive — access your files
  5. PostgreSQL — query your DB
  6. Filesystem — read/write local files
  7. Linear — manage backlog
  8. Sentry — analyze prod errors
  9. Stripe — query your payments
  10. Brave Search — web search

250+ public servers in total (May 2026 figure), plus all those companies build internally.

MCP vs alternatives: the verdict

Other approaches exist to connect AI and tools. Let's see why MCP wins.

MCP vs alternatives

 🔧Custom approach🔌MCP
Integration time1-5 days per tool30 min per tool
MaintenanceHigh (each API changes)Low (stable standard)
Multi-AI compatibilityNo (recoding)Yes (native)
MarketplaceNon-existent250+ public servers
SecurityDIYStandardized permissions
Unique specific use caseTotal controlSlight protocol constraint
CommunityNoneAnthropic + OpenAI + Google + 1000s devs

How to get started in 2026

Three paths depending on your profile:

1. You're a user (non-dev)

The simplest: use Claude Desktop (free). It integrates MCP natively. You install MCP servers in a few clicks from the interface, and Claude can suddenly access your Notion, your GitHub, etc.

Action: download Claude Desktop, go to Settings > MCP, add the Notion server. Ask Claude your first question about your docs.

2. You're a developer

Writing your own MCP server is simpler than you think. Anthropic provides SDKs in Python, TypeScript, Go, Rust.

📚Going deeper

Minimal example (Python):

from mcp.server import Server
from mcp.types import Tool

server = Server("my-server")

@server.tool()
async def calculate_vat(amount: float) -> str:
    vat = amount * 0.20
    return f"VAT = £{vat}, Total = £{amount + vat}"

server.run()

That's it. In 10 lines, you have an MCP server that Claude/ChatGPT/Gemini can call.

To go further:

  • Official repo: github.com/modelcontextprotocol
  • SDKs: Python, TypeScript, Go, Rust, Java
  • Docs: modelcontextprotocol.io
  • Example servers: 30+ open source implementations to copy

3. You're a business decision-maker

3 questions to ask your teams in May 2026:

  1. "Are our internal tools accessible via MCP?" — if not, your employees won't be able to use AIs effectively
  2. "Does our SaaS expose an MCP server?" — if not, your customers will ask for it
  3. "Do we have an MCP governance strategy?" — security, audit, permissions
Concrete use case: Shopify
Shopify published its MCP server in January 2026. Result: the 2 million merchants can now plug Claude/ChatGPT directly into their shop to generate product descriptions, analyze their sales, create campaigns. Adoption × 15 in 4 months on their AI APIs. A new ecosystem standard.

Limitations and blind spots

MCP isn't magic. Some important limitations to know:

What MCP does NOT solve
- Security remains your job: a poorly configured MCP server = open door to your data. Mandatory audit. - LLMs still hallucinate: even with MCP, a model can misinterpret data. Human validation necessary for critical actions (payments, deletion, email sends). - Latency: each MCP call adds time (50-300ms). For real-time responses, optimize. - The ecosystem evolves fast: the MCP spec changed 3 times in 18 months. Stay up to date.

Want to explore concretely?

🔌 Compare AIs that support MCP

See which AI models (Claude, GPT-5, Gemini...) natively support MCP in 2026, their prices, their strengths.

View comparison

The 2026 verdict

If you do enterprise AI, MCP is no longer optional. It's become the invisible infrastructure of all serious AI solutions.

Like HTTP for the web, like SQL for databases, MCP is becoming a universal standard. Early adopters have a competitive advantage. Latecomers will have to catch up painfully.

Practical recommendation:

  • This week: test Claude Desktop with 2-3 official MCP servers
  • This month: identify internal tools that would deserve an MCP server
  • This quarter: train someone to become your "MCP champion"
  • By end of 2026: your AI stack should be 100% MCP-native

Test your knowledge

🧠 Quiz
Question 1 of 5

Who created MCP?

Tags
MCPModel Context ProtocolAnthropicIntégration IAStandard

Read next