MCP Library updated 15 min read

Introduction to MCP: Connecting AI Assistants to Tools

What is Model Context Protocol (MCP)? Learn how this open protocol connects AI assistants to external tools and APIs.

Rajesh Praharaj

· Updated

Contents 13 sections
  1. TL;DR - What is MCP?
  2. The Problem: AI Assistants Are Isolated
  3. Enter MCP: The Universal Connector
  4. The MCP Ecosystem
  5. Real-World Use Cases
  6. Getting Started with MCP
  7. Security Best Practices
  8. MCP vs. Other Approaches
  9. Recent Major Updates (2025)
  10. Next Steps
  11. Quick Reference
  12. Summary
  13. Frequently Asked Questions

TL;DR - What is MCP?

MCP (Model Context Protocol) is an open JSON-RPC 2.0 protocol (modelcontextprotocol.io) that lets frontier AI models—including Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro (compare tool-calling latencies on our LLM Benchmarks hub)—connect to external tools, databases, and enterprise systems. Think of it as “USB-C for AI”: a universal standard governed by the Linux Foundation Agentic AI Foundation (AAIF).

Key Takeaways

  • Sub-10ms Local Execution & Streamable HTTP: Local stdio MCP servers execute tool calls with 2ms to 8ms IPC overhead and 50MB to 65MB RAM per process, while remote enterprise deployments use the 2025-11-25 Streamable HTTP transport with OAuth 2.1 + PKCE (S256) and RFC 9728 Protected Resource Metadata (/.well-known/oauth-protected-resource).
  • Context Window Economics: Registering a standard 10-to-15-tool MCP server consumes 1,200 to 2,500 tokens of JSON Schema definitions; uncurated 5+ server setups can burn 15,000 to 30,000 tokens (12% to 25% of a 128K context window) before the first user prompt unless mitigated with dynamic tool discovery.
  • Massive Open Ecosystem: Supported natively across Claude Desktop, Claude Code, ChatGPT, GitHub Copilot, Gemini, Cursor, and Windsurf, backed by 10,000+ community and official servers and 97 million+ monthly SDK downloads across @modelcontextprotocol/sdk and Python FastMCP.
  • Cross-Pillar Developer Workflow: Validate your claude_desktop_config.json payloads with our JSON Formatter & Validator and craft deterministic tool-calling system instructions using our AI Prompt Generator.
Quick Architecture Check: Do You Need Local stdio or Remote Streamable HTTP? (Click to Expand)
  • Choose Local stdio (npx / uvx / Docker) when connecting desktop IDEs (Claude Desktop, Cursor, VS Code) to local source code, local SQLite/PostgreSQL instances, or personal API tokens with 0ms network hop latency.
  • Choose Remote Streamable HTTP when deploying shared organizational tools behind an enterprise API gateway with centralized OAuth 2.1 token validation, RBAC audit logging, and horizontal auto-scaling.

Example: With GitHub MCP, you can say “Create a pull request with all the changes we discussed” and Claude will actually do it - not just explain how.


The Problem: AI Assistants Are Isolated

You’ve probably experienced this frustration:

Without MCP:

You: "Can you review the code in my project?"
AI: "I'd love to! Please paste the code you'd like me to review."
You: *copies 500 lines from file 1*
You: *copies 300 lines from file 2*  
You: *copies the package.json*
You: "Here's the code..."
AI: "Thanks! Now can you share the test files too?"
You: 😫

AI assistants are incredibly powerful, but they’re isolated bubbles. They can’t see your files, access your databases, or interact with your tools. Every conversation starts from scratch. For more on AI capabilities, see the AI Agents guide.

This limitation forces you to:

  • Copy-paste massive amounts of context
  • Describe systems instead of showing them
  • Manually execute every suggestion the AI makes
  • Re-explain your project in every new conversation

MCP solves this problem.


Enter MCP: The Universal Connector

Model Context Protocol (MCP) is an open protocol introduced by Anthropic in November 2024 and now maintained by the Linux Foundation’s Agentic AI Foundation (as of December 2025). It has become the industry standard for AI tool integration, adopted by all major AI providers including OpenAI, Microsoft, and Google. For a complete timeline of how MCP and other AI technologies evolved throughout 2025, see our AI in 2025: Year in Review.

The “USB for AI” Analogy

Before USB, every device had its own proprietary connector:

  • Printers needed printer cables
  • Cameras needed camera cables
  • Phones needed phone cables

USB created a universal standard. Now everything just works.

MCP does the same for AI:

  • Before MCP: Every AI tool had its own integration system (ChatGPT plugins, Claude tools, etc.)
  • After MCP: Any AI tool can use any MCP server
Architecture LayerRole in the Protocol StackConnected Examples & Transports
1. MCP Hosts (AI Applications)User-facing AI chat apps, IDEs, and autonomous agent harnessesClaude Desktop, Cursor, VS Code + Copilot/Continue, Windsurf, ChatGPT
2. MCP Protocol TransportBidirectional JSON-RPC 2.0 message bus handling tools/list, tools/call, and resources/readLocal stdio pipes (2ms–8ms latency) or Remote Streamable HTTP (POST/GET with SSE)
3. MCP Servers (Capability Providers)Isolated processes or microservices exposing domain-specific tools, resources, and promptsGitHub Server, PostgreSQL Server, Filesystem Server, Notion Server, Slack Server

How It Works

MCP has three main components:

ComponentWhat It IsExamples
MCP HostsAI applications that use MCPClaude Desktop, Cursor, Windsurf
MCP ClientsProtocol handlers inside hostsBuilt into the apps above
MCP ServersPrograms that expose capabilitiesGitHub, Postgres, Filesystem servers

When you configure an MCP server:

  1. Your AI tool (host) starts the MCP server
  2. The server tells the AI what tools are available
  3. The AI can now use those tools during conversations

The MCP Ecosystem

The MCP ecosystem has exploded since its launch, reaching 10,000+ active servers by December 2025. Here’s what’s available:

Official Reference Servers (by Agentic AI Foundation)

These are maintained by the Linux Foundation’s Agentic AI Foundation:

ServerWhat It DoesStatus
FilesystemRead/write local filesCore
GitRepository operationsCore
PostgreSQLDatabase queriesCore
SQLiteLocal database accessCore
FetchWeb content retrievalCore
MemoryKnowledge graph storageCore
PuppeteerBrowser automationCore
TimeTimezone conversionsCore

Official Integrations (by Companies)

Major companies have built official MCP servers:

CompanyServerWhat It Does
MicrosoftGitHub MCPManage repos, issues, PRs
PlaywrightPlaywright MCPBrowser automation
AtlassianAtlassian MCPJira + Confluence
AWSAWS MCPCloud services
CloudflareCloudflare MCPWorkers, KV, R2
SentrySentry MCPError monitoring
MongoDBMongoDB MCPDocument database
NotionNotion MCPWorkspace management
SlackSlack MCPTeam messaging

Community Servers

The community has built 9,000+ additional servers for:

  • Google Drive, Dropbox, OneDrive
  • Airtable, Linear, Trello
  • Discord, Telegram, WhatsApp
  • Docker, Kubernetes
  • Stripe, Shopify
  • SAP, Oracle, Salesforce
  • Custom enterprise integrations
  • And thousands more…

MCP Clients (AI Tools That Support MCP)

ToolTypeMCP SupportAdded
Claude DesktopChat App✓ NativeNov 2024
ChatGPTChat App✓ NativeEarly 2025
Microsoft CopilotChat/IDE✓ NativeEarly 2025
GeminiChat App✓ NativeDec 2025
CursorAI IDE✓ NativeNov 2024
WindsurfAI IDE✓ Native2024
VS Code + ContinueIDE Extension✓ Via extension2024
ClineVS Code Extension✓ Native2024
ZedEditor✓ Native2024

Real-World Use Cases

Let’s look at practical examples of what you can do with MCP:

‍ Developer Workflow

MCP shines in development environments. For more on AI-powered coding, see the AI-Powered IDEs Comparison guide.

Without MCP:

You: "Review my authentication code"
AI: "Please paste the code..."
*You spend 5 minutes copying files*

With MCP (Filesystem + Git + GitHub):

You: "Review the authentication code in src/auth/, 
      check git history for recent changes, 
      and create an issue for any security concerns"

AI: *reads files from src/auth/*
    *checks git log for recent commits*
    *analyzes code*
    *creates GitHub issue #47: "Potential session fixation vulnerability"*

"I reviewed the authentication module. Found 2 issues:
1. Created GitHub Issue #47 for a session fixation risk
2. The password hashing uses an outdated algorithm

Here are my recommended fixes..."

Data Analysis

With MCP (PostgreSQL + Filesystem):

You: "What were our top 10 products last month? 
      Export to a CSV on my desktop."

AI: *queries database*
    *writes file to ~/Desktop/top_products.csv*

"Done! I found the top 10 products by revenue:
1. Widget Pro - $45,230
2. Gadget Plus - $38,100
...
Exported to your Desktop as 'top_products.csv'"

Productivity

With MCP (Notion + Slack):

You: "Summarize what happened in #engineering today 
      and add action items to my Notion task database"

AI: *reads Slack channel*
    *creates Notion tasks*

"Here's today's summary from #engineering:
- Deployed v2.3.1 to production
- Discussed new auth system

I added 3 action items to your Notion:
✅ Review PR #892
✅ Schedule security audit
✅ Update documentation"

Research

With MCP (Fetch + Memory):

You: "Research the latest trends in AI agents. 
      Remember what you find for future reference."

AI: *fetches articles from web*
    *stores in knowledge graph*

"I researched AI agent trends and saved key findings:
- Multi-agent systems are trending
- Tool use is the fastest growing capability
- MCP adoption is accelerating

I've stored this in memory. Ask me about AI agents 
anytime and I'll remember this context."

For more on research techniques with AI, see the RAG, Embeddings & Vector Databases guide.


Getting Started with MCP

Ready to try MCP? Here’s a quick start:

Step 1: Choose Your MCP Client

Pick an AI tool that supports MCP:

If You Want…Use This
General chat + toolsClaude Desktop
AI-powered codingCursor or Windsurf
VS Code integrationContinue or Cline extension

Step 2: Configure Your First MCP Server

Here’s how to add the Filesystem MCP server to Claude Desktop:

1. Find your config file:

  • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
  • Windows: %APPDATA%/Claude/claude_desktop_config.json
  • Linux: ~/.config/Claude/claude_desktop_config.json

2. Add this configuration:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/yourname/projects"
      ]
    }
  }
}

3. Restart Claude Desktop

4. Test it:

You: "What files are in my projects folder?"
Claude: "I can see the following files and folders..."

Step 3: Add More Servers

Once you’re comfortable, add more servers:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/projects"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_your_token_here"
      }
    },
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres", "postgresql://localhost/mydb"]
    }
  }
}

Next Article: How to Install and Configure MCP Servers - Complete setup guide for Claude, Cursor, VS Code, and more.


Security Best Practices

MCP is designed with security in mind, but you should still follow best practices:

✓ Do’s

PracticeWhy
Grant minimal accessOnly allow directories/resources you need
Use read-only when possibleFilesystem MCP supports read-only Docker bind mounts
Review server permissionsUnderstand what each server can do
Keep servers updatedRun npx to always get latest versions
Use official servers firstThey’re audited and maintained

✗ Don’ts

PracticeRisk
Don’t grant home directoryExposes all your files
Don’t share config filesMay contain API keys
Don’t use unverified serversCould be malicious
Don’t ignore permission promptsServers ask for reason

For more on responsible AI usage, see the Understanding AI Safety, Ethics, and Limitations guide.

How MCP Security Works

Security BoundaryPermitted OperationsBlocked by Default
Host Process (Claude Desktop / Cursor)Spawns local MCP server process and mediates user human-in-the-loop tool approval promptsUnapproved destructive tool execution without user consent
Sandboxed MCP Server ProcessAccesses only explicitly passed CLI directories (e.g., /Users/me/projects canonicalized via fs.realpathSync) or scoped env API tokensArbitrary parent directory traversal (../), symlink escapes (patched in CVE-2025-53109), and unmounted system files
Host OS & External NetworkOutbound HTTPS calls only from servers explicitly designed for API/web access (e.g., GitHub, Fetch)Unauthorized access to ~/.ssh, system keychain, or unconfigured local databases

Key security features:

  • Local execution: Servers run on your machine, not in the cloud
  • Sandboxed access: Only resources you explicitly allow
  • No network by default: Most servers don’t make external calls
  • Transparent operation: You can inspect server source code

MCP vs. Other Approaches

How does MCP compare to other AI integration approaches?

FeatureMCPLegacy PluginsCustom APIs
Universal✓ All major AI tools✗ Platform-specific✗ Build each
Local execution✓ Your machine✗ Cloud-onlyDepends
Open source✓ Linux Foundation✗ ProprietaryDepends
Privacy✓ Data stays local✗ Sent to cloudDepends
Customizable✓ Build your own✗ Limited✓ Full control
Ecosystem✓ 10,000+ serversDeprecatedN/A
Industry adoption✓ OpenAI/MS/GoogleLegacyN/A

Bottom line: MCP has become the industry standard, supported by all major AI providers. It gives you the flexibility of custom APIs with the ease of configuration, while keeping your data secure.


Recent Major Updates (2025)

MCP has evolved significantly since launch:

December 2025: Linux Foundation Donation

  • Agentic AI Foundation formed under Linux Foundation
  • Founding supporters: OpenAI, Microsoft, Google, AWS, Anthropic, Block, Cloudflare
  • Open governance: Ensures MCP remains neutral and community-driven
  • Mission: Prevent fragmentation, accelerate innovation in agentic AI

November 2025: Major Protocol Update

  • Task-based workflows: Track long-running operations
  • Streamlined OAuth: URL-based client registration
  • Asynchronous operations: Better handling of slow tools
  • 🆔 Server identity: Improved security and authentication
  • Official extensions: Standardized extension mechanism

New Features in Claude

  • 75+ MCP connectors in Claude’s directory
  • Tool Search: Find tools across many connected servers
  • Programmatic Tool Calling: Optimized for large-scale deployments

What’s Coming

  • Managed remote MCP servers (Google Cloud already offers this)
  • Visual configuration tools
  • Mobile app support
  • Enterprise SSO and audit logging

Next Steps

Ready to dive deeper? Here’s your learning path:

Beginner Path

  1. ✓ You are here: Introduction to MCP
  2. → Next: Install & Configure MCP Servers
  3. → Then: Filesystem MCP Guide
  4. → Then: GitHub MCP Guide

Developer Path

  1. ✓ You are here: Introduction to MCP
  2. → Next: Build Your First MCP Server
  3. → Then: PostgreSQL MCP Guide
  4. → Then: Best MCP Combinations

Power User Path

  1. ✓ You are here: Introduction to MCP
  2. → Next: GitHub MCP Guide
  3. → Then: Notion MCP Guide
  4. → Then: Memory MCP Guide

Quick Reference

Essential Commands

# Run any MCP server with npx (no install needed)
npx -y @modelcontextprotocol/server-filesystem /path/to/allow

# Test MCP servers with Inspector
npx @modelcontextprotocol/inspector

# Check Node.js version (v18+ required)
node --version

Config File Locations

PlatformPath
macOS~/Library/Application Support/Claude/claude_desktop_config.json
Windows%APPDATA%/Claude/claude_desktop_config.json
Linux~/.config/Claude/claude_desktop_config.json
ServerPackageUse Case
Filesystem@modelcontextprotocol/server-filesystemRead/write files
GitHub@modelcontextprotocol/server-githubManage repos, issues, PRs
PostgreSQL@modelcontextprotocol/server-postgresQuery databases
Playwright@playwright/mcpBrowser automation
Memory@modelcontextprotocol/server-memoryPersistent knowledge

Resources


Summary

MCP (Model Context Protocol) has become the industry standard for AI integrations:

  • ✓ Universal standard adopted by OpenAI, Microsoft, Google, and Anthropic
  • ✓ 10,000+ servers for every use case imaginable
  • ✓ Linux Foundation governance ensures openness and neutrality
  • ✓ Secure by design with local execution
  • ✓ Easy to use with simple JSON configuration
  • ✓ Free and open source forever

The days of copy-pasting context into AI chats are coming to an end. With MCP, your AI assistant becomes a true partner that can access your tools, data, and workflows directly.

Next: Learn how to Install and Configure MCP Servers →


Have questions about MCP? Found an issue with this guide? Contact us or join the MCP community on Discord.

With the core architecture of MCP explained in this foundational MCP tutorial, you are ready to install your first server and build multi-tool agent workflows.

Frequently Asked Questions

What is MCP (Model Context Protocol)?

MCP is an open protocol originally created by Anthropic and now maintained by the Linux Foundation's Agentic AI Foundation. It standardizes how AI assistants connect to external data sources and tools. Think of it as 'USB for AI' - a universal way for AI to access databases, APIs, file systems, and more securely.

Which AI tools support MCP?

MCP is now an industry standard supported by Claude, ChatGPT, Microsoft Copilot, Gemini, Cursor, VS Code, Windsurf, and many more. OpenAI, Microsoft, and Google all integrated MCP in 2025.

Is MCP free to use?

Yes! MCP is an open protocol and all official reference servers are free and open source. Some third-party MCP servers may require API keys for the services they connect to (like GitHub or Notion).

Do I need to be a developer to use MCP?

No! While developers can build custom MCP servers, anyone can use existing MCP servers by simply adding them to their AI tool's configuration. It's as easy as copy-pasting a few lines.

Is MCP secure?

MCP is designed with security in mind. Servers run locally on your machine, you explicitly grant permissions, and there's no cloud intermediary for most servers. You control exactly what data your AI can access.

What's the difference between MCP and plugins?

Unlike proprietary plugins tied to one platform, MCP is a universal standard. An MCP server works with any compatible AI tool - Claude, Cursor, VS Code, etc. Build once, use everywhere.

Can I use MCP with ChatGPT?

Yes! OpenAI integrated MCP support in early 2025. ChatGPT, Microsoft Copilot, and Google's Gemini all support MCP now. MCP has become the industry standard for AI tool integration.

How many MCP servers are available?

There are over 10,000 active MCP servers as of December 2025, including official servers maintained by the Agentic AI Foundation, integrations from companies like GitHub, AWS, and Atlassian, plus thousands of community-built servers. With MCP explained from JSON-RPC 2.0 primitives to Streamable HTTP transport in this MCP tutorial, you are ready to connect your first local and remote servers.

Written by Rajesh Praharaj

Product Leader & AI Builder

Product leader and AI builder focused on practical LLM systems, Model Context Protocol (MCP) integrations, agentic workflows, and free browser-based tools for product & engineering teams.

Was this page helpful?

Let us know if you found what you were looking for.