Introduction to Model Context Protocol (MCP)

2026 Status

MCP is now the industry standard for connecting AI agents to tools and data sources. All major platforms (Claude, ChatGPT, Gemini) support it. If you're building agents that need to access external tools, MCP is how you do it.

What is MCP? (2026 Definition)

MCP is the universal protocol for giving AI agents access to tools, databases, and external data.

Think of it like USB-C for AI:

• Before MCP: Every AI platform had its own custom way to connect tools
• With MCP: One standard protocol, works everywhere

Why MCP Matters in 2026

The Problem It Solved

2023-2024: Want Claude to access your database? Build custom integration. Want ChatGPT to read Slack? Different custom integration. Want Gemini to browse files? Another custom integration.

2026: Build one MCP server. Works with all AI platforms.

Real Example

Before MCP: every AI platform needs its own custom integration to reach Slack.

With MCP: build one server, and every platform reaches Slack through it.

One integration, works everywhere.

What MCP Actually Does

MCP servers expose capabilities to AI agents:

1. Tools (Functions the AI can call)

// MCP Server exposes tools
{
  name: "search_slack",
  description: "Search Slack messages",
  parameters: {
    query: "string",
    channel: "optional string"
  }
}

Agent uses it:

Agent: "Find mentions of 'budget' in Slack"
→ Calls: search_slack("budget")
→ Gets: [list of messages]
→ Responds to user with findings

2. Resources (Data the AI can read)

// MCP Server exposes resources
{
  uri: "slack://channels/general/messages",
  name: "General channel messages",
  mimeType: "application/json"
}

3. Prompts (Pre-packaged instructions)

// MCP Server provides prompts
{
  name: "analyze_thread",
  description: "Analyze a Slack thread for action items",
  arguments: { thread_url: "string" }
}

Popular MCP Servers (2026)

Anthropic Reference Servers

• filesystem - Read/write files
• memory - Persistent memory for agents
• git - Read, search, and manipulate Git repositories
• sequential-thinking - Structured, multi-step reasoning

info

By 2026 most of the original reference servers from Anthropic's modelcontextprotocol/servers repo — including GitHub, Postgres, Google Drive, and Slack — have been archived in favor of vendor-maintained servers. GitHub, for example, now maintains its own official server at github/github-mcp-server rather than the community @modelcontextprotocol/server-github package. When picking a server for Slack, Postgres, or Google Drive, look for one actively maintained by the vendor or a well-known community project rather than the original archived examples.

Community Servers

Find hundreds at modelcontextprotocol.io

Popular:

• Notion MCP
• Salesforce MCP
• Linear MCP
• Figma MCP
• AWS MCP
• Browser automation MCP

How to Use MCP (User Perspective)

In Claude Desktop

1. Install MCP server (via npm, pip, or binary)
2. Configure in Claude settings
3. Claude can now use those tools automatically

Example - Add GitHub access (via the local Docker-based server):

// In Claude config
{
  "mcpServers": {
    "github": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "GITHUB_PERSONAL_ACCESS_TOKEN", "ghcr.io/github/github-mcp-server"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "your-github-pat"
      }
    }
  }
}

GitHub also runs a hosted, remote MCP endpoint (api.githubcopilot.com/mcp/), but as of 2026 it's restricted to GitHub Copilot integrations (VS Code, JetBrains, and similar) and rejects connections from third-party MCP clients like Claude Desktop. For Claude Desktop, use the local Docker-based server shown above with a personal access token; the original community npm package has been archived in favor of GitHub's own server.

Now Claude can:

• Read your repos
• Create issues
• Comment on PRs
• Search code
• Check CI status

In ChatGPT

1. Native MCP support: connect an MCP server as a ChatGPT app in Settings, or via Developer Mode. Apps are discovered and enabled through ChatGPT's plugin directory, and the underlying connection still speaks MCP
2. Configure authentication
3. ChatGPT can now call those tools directly

Note: this is separate from Custom GPT "Actions," an older OpenAPI-based mechanism for adding tool calls to a Custom GPT. MCP didn't replace Actions - the two mechanisms coexist, but MCP is the one that works the same way across platforms.

In Gemini

1. Native MCP support: point the Gemini API or Gemini CLI directly at an MCP server endpoint
2. Configure authentication if the server requires it
3. Gemini calls the server's tools directly, alongside its own built-in function calling

Building Your First MCP Server

You should build an MCP server if:

• You have a tool/database AI agents need to access
• You want to use it across multiple AI platforms
• You need structured tool access (not just prompting)

Simple MCP server (Python):

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("my-tool-server")

@mcp.tool()
async def get_data(query: str) -> str:
    """Fetch data from my system"""
    # Your logic here
    return f"Data for {query}"

if __name__ == "__main__":
    mcp.run()

Then use it with Claude, ChatGPT, or Gemini.

MCP vs Other Approaches

Approach	When to Use
MCP	Production agents needing reliable tool access
Function calling	Simple, single-platform integrations
Legacy ChatGPT Plugins (2023, deprecated)	Don't use — MCP-based apps replaced this older system
Prompting alone	No tool access needed

Real-World MCP Use Cases

1. Engineering Agent

MCP Servers: GitHub, Linear, Slack Agent can: Check issues, write code, create PRs, update tickets

2. Sales Agent

MCP Servers: Salesforce, HubSpot, email Agent can: Research leads, update CRM, send outreach

3. Research Agent

MCP Servers: Web scraper, PDF reader, database Agent can: Search web, extract data, query knowledge base

4. Analytics Agent

MCP Servers: Postgres, Google Analytics, Mixpanel Agent can: Query databases, generate reports, visualize data

Getting Started with MCP

For users:

1. Try Claude Desktop with built-in MCP servers
2. Install community servers for your tools
3. Configure and let agents use them

For developers:

1. Read MCP documentation
2. Try example servers from Anthropic
3. Build server for your tool/data
4. Publish to MCP registry

Common Questions

Q: Do I need to learn MCP to use AI agents? A: No - just install and configure servers. Building them requires coding.

Q: Is MCP only for Claude? A: No - industry standard works with all major platforms.

Q: Can I connect to any API via MCP? A: Yes, if there's an MCP server for it, or you build one.

Q: Are MCP servers secure? A: They run locally or on your infrastructure, you control access.

MCP Timeline (2024–2026)

Late 2024: Anthropic announces and open-sources MCP Early 2025: Community builds first servers; ecosystem starts forming 2025: OpenAI and Google add support; MCP becomes the cross-platform standard December 2025: Governance of MCP moves to the newly formed Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation, with Anthropic joining as a co-founder and platinum member alongside AWS, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI — and continuing to provide MCP's lead core maintainer 2026: Established as the default way to connect AI to tools across Claude, ChatGPT, and Gemini

Why a Vendor-Neutral Standard Matters

Handing MCP's governance to the Agentic AI Foundation (AAIF) — a directed fund under the Linux Foundation — in December 2025 was a deliberate move. Open standards last longer when no single company controls their direction — think of how HTTP, USB, or Kubernetes outgrew their original stewards. With Anthropic as one platinum member among several rather than the sole owner, companies building MCP servers and clients can trust that the protocol will evolve based on broad industry needs rather than one vendor's roadmap. That's a big part of why MCP became the standard every major AI platform adopted, instead of staying a Claude-only feature.

What's Next?

You now understand what MCP is, why it exists, and what an MCP server exposes. In the next lesson, we'll dive deeper into the three building blocks every MCP server is made of:

1. Tools - Functions the AI can call to take action
2. Resources - Data the AI can read for context
3. Prompts - Reusable instruction templates the server provides

You'll also learn how a client like Claude Desktop discovers what a server offers before ever using it.

Ready to go deeper? Continue to: Understanding Context Components

Knowledge Check

Before moving on, make sure you can answer these questions:

1. What problem did MCP solve that custom, one-off integrations couldn't?
2. What are the three things an MCP server can expose to an AI agent?
3. What's the difference between a Tool and a Resource in MCP?
4. Why did MCP's governance move to the Agentic AI Foundation (under the Linux Foundation) in December 2025, and why does that matter?

---

Quick Reference

MCP in One Sentence: An open, vendor-neutral protocol that lets AI agents discover and use tools, data, and prompt templates from any MCP server, through one consistent connection.

Client-Server Architecture: Your AI platform (Claude, ChatGPT, Gemini) is the client. Each tool or data source you connect is exposed through an MCP server. One server works with every MCP-compatible client.

The Three Capabilities: Tools (callable functions), Resources (readable data), Prompts (reusable instruction templates).

Key Insight: Build the integration once as an MCP server, and every AI platform that speaks MCP can use it — no more rebuilding the same connector for each vendor.