AI Collaboration MCP Server

🚧 Work in Progress - Active Development 🚧

A Model Context Protocol (MCP) server designed to facilitate direct AI-to-AI collaboration between Claude and Gemini, eliminating the need for human intermediation in development workflows.

Note: This project is under active development. While core features are functional, some aspects are still being refined. Contributions and feedback are welcome!

🎯 Project Goal

Enable truly autonomous AI-to-AI collaboration where:

• AI agents work continuously on complex projects

• Human intervention is minimal (ideally just starting the process)

• Agents create comprehensive project plans and execute 100+ phases autonomously

• Work continues until project completion or critical blocker

Related MCP server: Gemini Collaboration MCP Server

🚀 Quick Start

Both AIs just run:

@ai-collab init {"agentName": "gemini", "autonomous": true}  // For Gemini (CTO)
@ai-collab init {"agentName": "claude", "autonomous": true}   // For Claude (Developer)

That's it! The init command:

• Loads existing project context and state

• Creates or resumes a comprehensive project plan

• Automatically detects and continues pending work

• Shows critical tickets and blockers

• Starts autonomous execution loops

🌟 Recent Enhancements

Workflow Optimization (v2.0) 🚀

• Task Dependencies: Define dependsOn relationships between tasks

• Batch Task Creation: CTO can create multiple tasks in one command

• Priority-Based Work: Tasks are automatically prioritized (high/medium/low)

• Continuous Developer Mode: No waiting between tasks - automatic progression

• Smart Task Status: available, blocked, in_progress, in_review, completed

• Dependency Resolution: Tasks automatically unblock when dependencies complete

Autonomous Loop System

• 120-second check intervals for more natural workflow pacing

• 500 iteration maximum for extended autonomous operation

• Continuous work mode - agents keep working until project completion

• Manual loop execution - requires human to run check commands (automation WIP)

Project Plan Management

• Auto-generated 6-phase plans from PROJECT_REQUIREMENTS.md

• Smart phase progression - automatically moves to next phase when complete

• Duplicate task detection - prevents recreating completed features

• Ad-hoc mission support - pause main plan for urgent tasks

Enhanced Validation

• Ticket vs Task distinction - prevents confusion between bug reports and work items

• Role-based instructions - clearer guidance for CTO vs Developer roles

• Workflow enforcement - ensures proper task creation and submission flow

⚠️ Current Limitations

Automation Challenges

• Manual loop execution required - AI agents can't schedule their own checks

• PATH configuration needed - Claude/Gemini commands must be accessible

• API quota limits - Gemini has daily request limits that may be exceeded

Addressed Issues ✅

• Single task queuing → Now supports batch task creation

• Developer idle time → Continuous work mode implemented

• No task dependencies → Full dependency system added

• Random task order → Priority-based scheduling active

Remaining Challenges

• Agents occasionally create duplicate tasks (improved but not eliminated)

• Edit button functionality may need manual verification

• Loop execution still requires human intervention

Workarounds Available

• Automation scripts provided (mcp-automator.js) but require setup

• Manual loop execution instructions included

• Simulation mode for tracking when automation fails

Features

Core Capabilities

• Comprehensive Project Plans: 100+ phase autonomous execution capability

• One-Command Startup: Just init with autonomous flag

• Role-Based System: CTO, Developer, PM, QA, Architect roles

• Smart Task Management: Duplicate detection and phase progression

• Ticketing System: Track bugs, enhancements, tech debt

• Context Retention: Maintains state across sessions

• Mission Management: High-level objectives with auto-decomposition

• Code Review Workflow: Submit, review, and revision cycles

• Question & Answer System: Asynchronous clarifications

• Comprehensive Logging: Full audit trail

🆕 Enhanced Workflow Features (v2.0)

• Task Dependencies: Tasks can depend on other tasks with automatic blocking/unblocking

• Priority-Based Scheduling: High/medium/low priority with smart task selection

• Batch Task Creation: CTO can queue multiple tasks at once for efficiency

• Continuous Work Mode: Developer automatically moves to next available task

• Smart Status System: available, blocked, in_progress, in_review, completed

• Dependency Visualization: Clear indication of task dependencies and blockers

Installation

1. Clone this repository:

git clone https://github.com/yourusername/ai-collab-mcp.git
cd ai-collab-mcp

2. Install dependencies:

npm install

3. Make the server executable:

chmod +x src/index.js

Configuration

For Claude Code

Create a .mcp.json file in your project root:

{
  "mcpServers": {
    "ai-collab": {
      "command": "node",
      "args": [".mcp-server/src/index.js"],
      "cwd": "/path/to/your/project"
    }
  }
}

For Gemini

Configure in ~/.gemini/settings.json:

{
  "mcpServers": {
    "ai-collab": {
      "command": "node",
      "args": ["/path/to/ai-collab-mcp/src/index.js"]
    }
  }
}

Note: Gemini may require explicit instructions to execute MCP commands.

Usage

🎯 Autonomous Mode (Recommended)

Start with autonomous flag for continuous operation:

# Terminal 1 - Claude (Developer)
@ai-collab init {"agentName": "claude", "autonomous": true}

# Terminal 2 - Gemini (CTO)  
@ai-collab init {"agentName": "gemini", "autonomous": true}

# Terminal 3 - Manual Loop Execution (Required)
# Every 120 seconds, run:
@ai-collab get_loop_status {"agentName": "claude"}
@ai-collab get_loop_status {"agentName": "gemini"}

Automation Helpers (Experimental)

For reduced manual intervention:

# Run automation script (requires setup)
cd /path/to/project
node mcp-automator.js auto

# Or simulation mode (shows what would happen)
node mcp-automator-v2.js auto

See AUTOMATION.md for setup details.

Traditional Commands

CTO Tools

• send_directive - Create development tasks (now with dependencies & priority)

• send_batch_directives - Create multiple tasks at once

• review_work - Review submissions

• create_project_plan - Start comprehensive plan

• update_plan_progress - Move to next phase

Developer Tools

• get_all_tasks - View assigned work (sorted by priority)

• submit_work - Submit completed tasks

• ask_question - Request clarification

🆕 Enhanced Workflow Examples

Creating Tasks with Dependencies

// Single task with dependency
@ai-collab send_directive {
  "taskId": "KAN-002",
  "title": "Create database tables",
  "specification": "Create user and project tables",
  "priority": "high",
  "dependsOn": ["KAN-001"]  // Won't be available until KAN-001 is approved
}

// Batch creation with dependencies
@ai-collab send_batch_directives {
  "tasks": [
    {
      "taskId": "KAN-003",
      "title": "Setup database connection",
      "specification": "Configure PostgreSQL connection",
      "priority": "high"
    },
    {
      "taskId": "KAN-004", 
      "title": "Create user model",
      "specification": "Implement User model with validation",
      "priority": "medium",
      "dependsOn": ["KAN-003"]
    },
    {
      "taskId": "KAN-005",
      "title": "Create auth endpoints",
      "specification": "Implement login/register endpoints",
      "priority": "medium",
      "dependsOn": ["KAN-004"]
    }
  ]
}

Continuous Work Mode (Developer)

When the developer runs get_loop_status, they will:

1. See prioritized available tasks

2. Automatically start on the highest priority task

3. After submitting, immediately move to next task

4. Continue until all available tasks are complete

No more waiting between tasks! The developer keeps working continuously.

Project Plan Workflow

1. Automatic Plan Creation: On first init, generates 6-phase plan from requirements

2. Phase Progression: Automatically advances when all phase tasks complete

3. Duplicate Prevention: Skips tasks that match completed work

4. Ad-hoc Missions: Can pause main plan for urgent work

Example phases:

• Foundation & Basic Structure

• Core Interactive Features

• UI/UX Enhancement

• Data Persistence

• Advanced Features

• Polish & Quality Assurance

Data Storage

data/
├── tasks.json              # Task tracking
├── missions.json           # Active missions
├── project-state.json      # Project configuration
├── project-plans.json      # Comprehensive plans (NEW)
├── loop-states.json        # Autonomous loop tracking (NEW)
└── tickets/
    └── tickets.json        # Bug/enhancement tracking

Troubleshooting

Gemini Not Executing Commands

• Prefix with: "Execute the following MCP command:"

• Or: "Use the ai-collab tool to run:"

Duplicate Task Creation

• System now detects similar task names

• Manually clean duplicates from data/tasks.json if needed

Loop Not Continuing

• Ensure 120-second intervals between checks

• Verify agent hasn't exceeded maxIterations (500)

• Check API quotas haven't been exceeded

Contributing

This project needs help with:

• True automation (removing manual loop execution)

• Better Gemini CLI integration

• Improved duplicate detection algorithms

• Cross-platform automation scripts

1. Fork the repository

2. Create feature branch (git checkout -b feature/improvement)

3. Commit changes (git commit -m 'Add improvement')

4. Push branch (git push origin feature/improvement)

5. Open Pull Request

Roadmap

• Native scheduling in MCP server

• WebSocket/SSE for real-time updates

• Improved role switching

• Better error recovery

• Multi-project support

• Visual progress dashboard

License

MIT License - see LICENSE file for details.

Support

For issues, questions, or contributions, please open an issue on GitHub.

Remember: This is an experimental project pushing the boundaries of AI collaboration. Expect rough edges but exciting possibilities!