Tool-Specialized Agent Switching in Cursor - Dynamic MCP Tools Management

Feature Requests
1

Overview

Add support for multiple agent presets in Cursor, each with their own dedicated set of MCP tools, allowing users to switch between them manually or have them automatically selected based on context.

Current Limitation

Currently, Cursor loads all MCP tools defined in .cursor/mcp.json globally, which can lead to:

  • Tool namespace conflicts
  • Increased initialization time
  • Cognitive overhead when dealing with many tools
  • Limited context specialization

Also there’s a limit of 40 active tools loaded to the context.

Proposed Solution

Introduce agent presets that can be:

  1. Manually selected through the chat interface
  2. Automatically chosen by a local model based on context

Key Features

1. Agent Preset Configuration

{
  "agents": {
    "web-dev": {
      "name": "Web Development Agent",
      "description": "Specialized in web development tasks",
      "model": {
        "name": "claude-3-sonnet",
        "context_length": 200000,
        "description": "Optimized for code understanding and web development"
      },
      "mcpServers": {
        "browser-tools": { ... },
        "playwright": { ... },
        "strapi-graphql": { ... }
      }
    },
    "data-science": {
      "name": "Data Science Agent",
      "description": "Specialized in data analysis and ML",
      "model": {
        "name": "claude-3-opus",
        "context_length": 300000,
        "description": "Enhanced capabilities for mathematical and statistical tasks"
      },
      "mcpServers": {
        "jupyter": { ... },
        "pandas-tools": { ... },
        "sklearn-tools": { ... }
      }
    },
    "system-admin": {
      "name": "DevOps Agent",
      "description": "Specialized in system administration and DevOps",
      "model": {
        "name": "gpt-4",
        "context_length": 128000,
        "description": "Strong reasoning for system architecture and operations"
      },
      "mcpServers": {
        "kubernetes": { ... },
        "docker-tools": { ... },
        "aws-tools": { ... }
      }
    }
  }
}

2. Context-Aware Switching

  • Local model analyzes:
    • Current file types and content
    • Recent conversations
    • Active workspace
    • User’s explicit hints
  • Automatically suggests or switches to the most appropriate agent

3. UI Integration

  • Agent selector in chat window
  • Visual indicator of current agent
  • Quick switch shortcuts
  • Agent capabilities overview

Benefits

  1. Improved Performance

    • Load only necessary tools
    • Reduced memory footprint
    • Faster initialization

  2. Better Context Handling

    • Specialized tools for specific tasks
    • More relevant suggestions
    • Reduced tool confusion

  3. Enhanced User Experience

    • Clear indication of available capabilities
    • Seamless switching between contexts
    • Reduced cognitive load

  4. Extensibility

    • Community-shared agent presets
    • Custom agent configurations
    • Domain-specific optimizations

Implementation Considerations

  1. Backward Compatibility

    • Support existing .cursor/mcp.json format
    • Gradual migration path
    • Default agent for legacy setups

  2. Performance

    • Lazy loading of MCP servers
    • Efficient agent switching
    • Local model resource usage

  3. User Control

    • Override automatic selection
    • Custom agent creation
    • Preference persistence

Example Use Cases

  1. Full-Stack Development

    User: "Need to debug this API endpoint"
System: *Switches to Backend Agent with API testing tools*

User: "Now let's style the frontend"
System: *Switches to Frontend Agent with browser tools*

  2. Data Analysis Project

    User: "Let's analyze this dataset"
System: *Switches to Data Science Agent*

User: "Create a web dashboard for the results"
System: *Switches to Web Development Agent*

Next Steps

  1. Community feedback on:

    • Agent preset structure
    • Switching mechanics
    • UI/UX preferences

  2. Technical exploration:

    • MCP server management
    • Local model requirements
    • Performance impact

  3. Prototype development:

    • Basic agent switching
    • UI components
    • Configuration format