How Cursor, Claude Code, and Windsurf Use MCP for Agentic Coding

Developers using AI coding assistants are reporting 55% faster task completion after connecting their tools to external services via MCP. Yet most developers are still copy-pasting data between tabs instead of letting their agent reach out and grab it directly.

That changes today. This guide breaks down exactly how the three leading agentic coding tools - Cursor, Claude Code, and Windsurf - implement the Model Context Protocol.

---

What Is MCP (Model Context Protocol)?

MCP is an open standard that lets AI models talk to external tools using a single, unified protocol. Build one MCP server and every compatible client can use it.

Anthropic launched MCP on November 25, 2024. The latest spec version is 2025-06-18. Open-source, JSON-RPC 2.0 based, supported by Cursor, Claude Code, Windsurf, VS Code Copilot, and more.

The Problem MCP Solves

Before MCP: M AI models × N tools = M×N custom integrations. MCP collapses that to M+N.

Think of it as "USB-C for AI" - one standard plug. (For how the host, client, and server fit together, see the MCP architecture.)

MCP's Three Core Primitives

• Tools - Executable functions (e.g., create_github_issue, query_postgres)
• Resources - Read-only data (files, database schemas, API responses)
• Prompts - Templated workflows for multi-step tasks

Transport Types

• STDIO - Runs locally as a subprocess
• HTTP with SSE - Connects to remote servers over the network

---

What Is Agentic Coding?

Agentic coding is when an AI autonomously plans and executes multi-step coding tasks without you directing every move. Not autocomplete. Not a chatbot. A system that reads your codebase, reasons, and acts.

What a real agent can do with the right MCP setup:

1. Read a Jira ticket describing a new feature
2. Explore the relevant files in your codebase
3. Write the implementation across multiple files
4. Run the test suite and fix failures
5. Open a pull request on GitHub

Autocomplete vs. Agentic Coding

Feature	Autocomplete	Agentic Coding
Scope	Single line or block	Entire codebase
Reasoning	None	Multi-step planning
Tool use	No	Yes (via MCP)
Memory	No	Session-level context
Autonomy	Zero	High

---

How Cursor Uses MCP

Cursor supports MCP natively in its Agent and Composer modes, letting you connect to GitHub, Jira, Datadog, and more via a simple JSON config file.

Cursor MCP Config

Cursor reads MCP config from two locations:

• ~/.cursor/mcp.json - Global
• .cursor/mcp.json - Project-specific (commit to share with team)

Both support STDIO and HTTP transports.

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "your_token_here"
      }
    }
  }
}

Cursor's Agent Mode

Cursor's Composer / Agent Mode with MCP can:

• Read 15–20 files to understand context
• Edit multiple files simultaneously
• Run terminal commands and react to output
• Call MCP tools for external data

Cursor's latest release introduced a 3 Agents Window - orchestrate up to 8 parallel agents simultaneously.

Pre-Built MCP Integrations in Cursor

• Asana - Read and update tasks
• GitHub - Issues, PRs, code review
• Jira - Tickets and sprint data
• Datadog - Logs and metrics
• Browser tools - Web scraping and automation
• Sequential thinking - Multi-step reasoning chains

---

How Claude Code Uses MCP

Claude Code is Anthropic's CLI-based agentic coding tool, with the most flexible and powerful MCP implementation - supporting four transport types and three configuration scopes.

Claude Code isn't Claude the chatbot. It's a separate terminal tool that runs Claude models to execute real coding tasks.

Claude Code MCP Config

• .mcp.json - Project-scoped, committed to git
• ~/.claude.json - User-scoped, across all projects

CLI Commands

# Add an MCP server
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

# List configured servers
claude mcp list

# Remove a server
claude mcp remove sentry

Four Transport Types

• HTTP (streamable-http) - Recommended for remote
• SSE - Deprecated; use HTTP
• STDIO - Local processes
• WebSocket - Persistent bidirectional connections

Real Examples

Sentry error monitoring:

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

Then ask: "What are the most common errors in the last 24 hours?"

GitHub code review:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
  --header "Authorization: Bearer YOUR_GITHUB_PAT"

Then ask: "Review PR #456 and suggest improvements."

PostgreSQL queries:

claude mcp add --transport stdio db -- npx -y @bytebase/dbhub \
  --dsn "postgresql://readonly:pass@prod.db.com:5432/analytics"

Then ask: "What's our total revenue this month?" (For a deeper walkthrough of connecting Claude Code to internal tools, see our dedicated guide.)

MCP Resources via @ Mentions

Can you analyze @github:issue://123 and suggest a fix?

Reliability Features

• OAuth 2.0 for remote servers
• Auto-reconnect with exponential backoff (1s to 32s, 5 attempts)
• Dynamic tool updates via list_changed notifications

---

How Windsurf Uses MCP

Windsurf has the most beginner-friendly MCP experience with a visual marketplace UI that lets you install MCP servers without touching a config file.

Windsurf's agentic AI is called Cascade.

Windsurf MCP Setup

Two options:

1. Visual UI - Settings → MCP Servers, browse marketplace, click Install
2. Config file - Edit JSON directly

Supported Transports

STDIO and SSE.

Popular MCP Servers in Windsurf's Marketplace

• GitHub - Issues, PRs, code review
• Slack - Read and post messages
• Notion - Read and update docs
• Browser automation
• File system - Advanced file operations

---

Side-by-Side Comparison

Feature	Cursor	Claude Code	Windsurf
MCP Config Method	JSON file	CLI + JSON file	UI + JSON file
Transports Supported	STDIO, HTTP	STDIO, HTTP, SSE, WebSocket	STDIO, SSE
Agent Mode Name	Composer / Agent Mode	Claude Code (CLI)	Cascade
Parallel Agents	Up to 8	N/A	N/A
MCP Marketplace	No	No	Yes (visual UI)
Best For	Power users, IDE-native	Terminal/CLI workflows	Beginners, visual setup
Shared Config (team)	.cursor/mcp.json via git	.mcp.json via git	Shared workspace settings

Bottom line: Claude Code for terminal users. Windsurf's marketplace for smoothest onboarding. Cursor for IDE-native power with parallel agents.

---

Set Up Your First MCP Server (GitHub Example)

In Cursor

Step 1: Create .cursor/mcp.json in your project root.

Step 2: Add the GitHub MCP server config:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "your_token_here"
      }
    }
  }
}

Step 3: Restart Cursor.

Step 4: Open Agent Mode and type: "List my open GitHub issues."

In Claude Code

Step 1: Run in terminal:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
  --header "Authorization: Bearer YOUR_GITHUB_PAT"

Step 2: Verify:

claude mcp list

Step 3: Start a session and say: "Review PR #456 and suggest improvements."

In Windsurf

Step 1: Open Settings → MCP Servers.

Step 2: Search for "GitHub" in the marketplace.

Step 3: Click Install and enter your GitHub PAT.

Step 4: Open Cascade and say: "Show me my open pull requests."

---

Real-World Agentic Coding Workflows with MCP

Workflow 1: The Full Dev Loop

Tools: Cursor + GitHub MCP + Jira MCP

Agent reads a Jira ticket describing a bug. Explores relevant files, writes the fix, runs the test suite, opens a GitHub PR.

Workflow 2: The Data-Driven Debug

Tools: Claude Code + Sentry MCP + PostgreSQL MCP

Ask Claude Code: "What errors happened in the last 24 hours and which users were affected?" The agent queries Sentry, hits PostgreSQL, returns a formatted report.

Workflow 3: The Design-to-Code Pipeline

Tools: Windsurf + Figma MCP + Slack MCP

A designer updates a Figma component. Cascade reads the new design via Figma MCP, updates the React component, posts a Slack message with the diff.

---

Limitations and Things to Watch Out For

Security matters. MCP servers run locally with access to your tokens and environment. Don't commit API keys to .mcp.json.

Not all MCP servers are production-ready. Check GitHub stars, last commit date, and open issues before installing. (When no existing server fits, you can always build your own MCP server.)

Agentic loops can go wrong. Always review agent actions before committing. Use Cursor's approval gates and Claude Code's checkpoint system.

Context window limits still apply. Very large codebases may need chunking.

SSE transport is deprecated in Claude Code. Use HTTP instead.

Windsurf's MCP marketplace is growing but not exhaustive. Niche integrations still need manual JSON configuration.

---

FAQ

What is MCP in coding?

MCP (Model Context Protocol) is an open standard from Anthropic that lets AI coding tools connect to external services like GitHub, Jira, and databases using a single, unified protocol.

Does Cursor support MCP?

Yes. Cursor supports MCP via a JSON config file and uses it in Agent/Composer Mode to connect to tools like GitHub, Jira, and Datadog. Supports STDIO and HTTP transports.

Is Claude Code the same as Claude?

No. Claude is Anthropic's AI model (the chatbot). Claude Code is a separate CLI tool that uses Claude models to perform agentic coding tasks in your terminal.

Which tool has the easiest MCP setup?

Windsurf - thanks to its visual MCP marketplace. Claude Code is most powerful for CLI users. Cursor sits in the middle.

Can I use MCP with multiple tools at once?

Yes. You can run multiple MCP servers simultaneously in all three tools.

Is MCP free to use?

MCP itself is free and open-source (Apache 2.0 licensed). Cursor, Claude Code, and Windsurf all have free tiers and paid plans around $20/month.

What is agentic coding?

When an AI autonomously plans and executes multi-step coding tasks - reading files, writing code, running tests, using external tools - without you directing every step.

---

Useful Sources

• Anthropic MCP Announcement
• MCP Official Documentation
• Cursor MCP Docs
• Claude Code MCP Docs
• Windsurf MCP Docs
• MCP GitHub Repository