• AI-powered code editor built on VSCode
• Using Claude 3.7 Sonnet as base LLM
• Advanced Tool calling capabilities that enable agents to execute tasks in one go
• Native support of Model Context Protocol for extended functionality

Cursor editor interface

• Agent Mode - Completes coding tasks end-to-end
• Chat - AI-powered chat that understands your codebase
• Tab - Smart code autocomplete with multi-line predictions
• Ctrl+K - Edit and generate code with AI assistance

Cursor features overview

• Autonomous AI coding agent that independently explores, plans, and executes complex codebase changes
• Follows a systematic workflow:
  • 🧠 Understand your request and codebase context
  • 🔍 Explore relevant files and documentation
  • 📝 Plan changes with a step-by-step approach
  • 🔧 Execute necessary modifications across files
  • ✅ Verify results and fix any issues
• Creates checkpoints before making changes, allowing you to revert if needed

• Local-first scan – builds a Merkle-tree of your files (respects .gitignore/.cursorignore)
• Incremental updates – every 10 minutes it diff-checks and only uploads changed chunks
• Vector embeddings – files split into overlapping snippets and embedded server-side (with obfuscated paths)
• On-demand retrieval – queries return only chunk IDs; Cursor reads the matching lines locally before prompting the LLM

What Are Cursor Rules?

• Custom instructions for the AI within Cursor
• Override default behavior for specific use cases
• Can be global or project-specific
• Stored in .cursor/rules.json files

Best Practices

• Keep rules clear and specific
• Target only necessary modes
• Use project-specific rules for unique requirements
• Global rules for personal preferences

• Core Identity: "You are a powerful agentic AI coding assistant, powered by Claude 3.7 Sonnet"
• Primary Goal: Follow user instructions at each message
• Tool Integration: Access to file system, search and code manipulation tools
• Context Awareness: Automatically attaches information about current state
• Proactiveness: Balances task completion with user control

Cursor operates with a sophisticated system prompt

Tool Integration Guidelines

• Never refer to tool names when speaking to users
• Follow schema exactly and provide all parameters
• Explain intentions before calling tools

Code Change Rules

• Never output code directly to users
• Always read file contents before editing
• Group edits to the same file together
• Limit to 3 attempts for fixing linter errors

Code Format Standards

• Create dependency files when building from scratch
• Never generate long hashes or binary code
• Fix reasonable linter errors

Search Optimization

• Prefer semantic search over grep/file search
• Read larger sections at once
• Stop searching once sufficient information found

• Open protocol that standardizes how applications provide context to LLMs
• Acts like a USB-C port for AI applications
• Connects AI models to data sources and tools
• Enables switching between LLM providers
• Ensures data security within your infrastructure

flowchart LR
  %% Define a reusable class for consistent styling
  classDef common fill:#e0e0ff,stroke:#4b0082,stroke-width:1px,color:#000080;

  subgraph "Your Computer"
    direction TB
    client["Host with MCP Client
(Claude, IDEs, Tools)"]:::common

    subgraph "MCP Servers"
      direction LR
      A["MCP Server A"]:::common --> DataA[(Local Data Source A)]:::common
      B["MCP Server B"]:::common --> DataB[(Local Data Source B)]:::common
      C["MCP Server C"]:::common --> Web[(Web APIs)]:::common
    end
  end

  subgraph Internet
    Remote[(Remote Service C)]:::common
  end

  client -->|MCP Protocol| A
  client -->|MCP Protocol| B
  client -->|MCP Protocol| C

  DataA --> A
  DataB --> B
  Web   --> Remote
									

MCP architecture overview

• MCP Hosts: Programs like Cursor, Claude Desktop, or AI tools
• MCP Clients: Protocol clients with 1:1 server connections
• MCP Servers: Lightweight programs exposing capabilities
• Data Sources: Local files, databases, services
• Remote Services: External APIs and web services

flowchart TD
  classDef hosts fill:#f9f9ff,stroke:#333,stroke-width:1px
  classDef clients fill:#e6f3ff,stroke:#333,stroke-width:1px
  classDef servers fill:#f0fff0,stroke:#333,stroke-width:1px
  classDef data fill:#fff0f0,stroke:#333,stroke-width:1px
  classDef remote fill:#f5f5f5,stroke:#333,stroke-width:1px
  
  Host["MCP Hosts
(Cursor, Claude Desktop)"]:::hosts
  Client["MCP Clients
(Protocol clients)"]:::clients
  Server["MCP Servers
(Lightweight programs)"]:::servers
  Data["Data Sources
(Files, DBs, Services)"]:::data
  Remote["Remote Services
(External APIs)"]:::remote
  
  Host --> Client
  Client --> Server
  Server --> Data
  Server --> Remote
									

MCP Architecture Diagram

Configuration Locations:

• Project Configuration:
  .cursor/mcp.json in project directory
• Global Configuration:
  ~/.cursor/mcp.json in home directory

Configuration Structure:

• command: The executable to run
• args: Command line arguments
• env: Environment variables

// Example Project-specific MCP configuration
{
  "mcpServers": {
    "database-accessor": {
      "command": "python",
      "args": ["-m", "mcp_database_server"],
      "env": {
        "DB_CONNECTION": "postgres://user:pass@localhost/db",
        "API_KEY": "sk-xxxx"
      }
    },
    "documentation-lookup": {
      "command": "npx",
      "args": ["-y", "mcp-docs-server"],
      "env": {
        "DOCS_PATH": "./documentation"
      }
    }
  }
}

Automatic Tool Usage:

• Composer Agent automatically uses relevant MCP tools
• Tools are listed under 'Available Tools' in MCP settings
• You can intentionally prompt tool usage by mentioning the tool

MCP Tool usage in Cursor