title: Yuga Planner emoji: 🐍 colorFrom: purple colorTo: gray sdk: docker app_port: 7860 license: apache-2.0 tags: ["agent-demo-track"]

Yuga Planner 🐍

This project was developed for the Hugging Face Agents MCP Hackathon!

Yuga Planner is a neuro-symbolic system prototype: it provides an agent-powered team scheduling and task allocation platform built on Gradio.

It takes a project description, breaks it down into actionable tasks through a LLamaIndex agent, then uses Timefold to generate optimal employee schedules for complex projects.

🚀 Try It Now

Live Demo: https://huggingface.co/spaces/blackopsrepl/yuga-planner

Source Code on GitHub: https://github.com/blackopsrepl/yuga-planner

MCP Chat Interface Usage

1. Navigate to the Chat tab in the Gradio interface
2. Upload your calendar file (.ics) to provide existing commitments (optional)
3. Type your scheduling request using natural language:
   hljs language-css

   Copy

   "Create a new EC2 instance on AWS"
   "Create a Svelte UI that allows me to query a postgresql database"
   "Develop a chatbot UI based on Gradio"
   

4. Receive intelligent responses that combine conversational AI with task scheduling capabilities
5. View formatted schedules with rich table output and status indicators

MCP Tool Usage

1. In any MCP-compatible chatbot or agent platform:
   hljs language-less

   Copy

   use yuga-planner mcp tool
   Task Description: [Your task description]
   

2. Attach your calendar file (.ics) to provide existing commitments
3. Receive optimized schedule that integrates your new task with existing calendar events

Architecture

Yuga Planner follows a service-oriented architecture with clear separation of concerns:

Core Services Layer

• DataService: Handles data loading, processing, and format conversion from various sources (Markdown, calendars)
• ScheduleService: Orchestrates schedule generation, solver management, and solution polling
• StateService: Centralized state management for job tracking and schedule storage
• LoggingService: Real-time log streaming for UI feedback and debugging
• MockProjectService: Provides sample project data for testing and demos
• MCPClientWrapper: Manages MCP server functionality and tool definitions

MCP Integration Components

• ToolCallAssembler: Processes streaming tool call deltas from Nebius API into complete tool calls
• ToolCallProcessor: Executes completed tool calls via MCP backend with JSON repair functionality
• Chat Interface: Unified conversational AI + task scheduling with intelligent tool detection
• Streaming Handler: Real-time response processing with progress indicators and error recovery

System Components

• Gradio UI: Modern web interface with real-time updates and interactive schedule visualization
• Task Composer Agent: Uses LLamaIndex + Nebius AI for intelligent task decomposition and estimation
• Constraint Solver: Timefold optimization engine for optimal task-to-employee assignments
• MCP Integration: Model Context Protocol endpoint for agent workflow integration

🌟 Key Features

🎯 Two Usage Modes

Yuga Planner operates as two separate systems serving different use cases:

1. 💬 MCP Chat Interface

Purpose: Conversational AI with integrated task scheduling capabilities

• Access: Chat tab in the Gradio interface
• Input: Natural language requests + optional .ics calendar files
• Features:
  • Intelligent tool detection based on scheduling keywords
  • Streaming responses with real-time tool call assembly
  • Rich table formatting for schedule results
  • Dual response system (Nebius API + MCP fallback)
• Use Case: Interactive scheduling through natural conversation

2. 🤖 MCP Personal Tool

Purpose: Individual task scheduling integrated with personal calendars

• Access: Through MCP-compatible chatbots and agent platforms
• Input: Attach .ics calendar files + natural language task descriptions
• Team: Schedules against your personal calendar and existing commitments
• Use Case: Personal productivity and task planning around existing appointments

Example MCP Usage:

User: use yuga-planner mcp tool
Task Description: Create a new EC2 instance on AWS
[Attaches calendar.ics file]

Tool Response: Optimized schedule created - EC2 setup task assigned to
available time slots around your existing meetings

🧩 MCP Chat Interface Technical Details

Intelligent Tool Detection

The chat interface automatically detects scheduling requests using keyword analysis:

scheduling_keywords = [
    'schedule', 'task', 'calendar', 'plan', 'organize',
    'meeting', 'appointment', 'project', 'deadline',
    'create', 'setup', 'implement', 'develop'
]

Streaming Tool Call Processing

• Delta Assembly: Collects 200+ streaming deltas into complete tool calls
• JSON Repair: Handles malformed JSON from streaming responses
• Progress Indicators: Real-time feedback during tool processing
• Timeout Protection: 60-second timeout for MCP operations

Dual Response System

• Primary: Nebius API with tool calling capabilities
• Fallback: Direct MCP backend invocation when tool assembly fails
• Error Recovery: Comprehensive error handling and graceful degradation

Features:

• Accepts calendar files and user task descriptions via chat interface
• Parses existing calendar events and new task requirements
• Full schedule solving support - generates optimized task assignments
• Returns complete solved schedules integrated with personal calendar
• Designed for seamless chatbot and agent workflow integration

Current Limitations:

• Multi-timezone support: Currently operates in a single timezone context with UTC conversion for consistency. Calendar events from different timezones are normalized to the same scheduling context.

See the CHANGELOG.md for details on recent MCP-related changes.

Work in Progress

• 🔧 Gradio UI overhaul: Enhanced user experience and visual improvements
• 🔍 Migration to Pydantic models: Type-safe data validation and serialization
• ⚡ Enhanced timezone support: Multi-timezone calendar integration for international scheduling

Future Work

System Integration Roadmap

• Unified scheduling engine that can handle both team management and personal productivity in one interface
• Hybrid workflows where personal tasks can be coordinated with team projects
• Cross-system data sharing between web demo projects and personal MCP calendars
• Seamless switching between team management and individual task planning modes

Core Feature Enhancements

• Multi-Tool Support: Extend chat interface to support additional MCP tools beyond scheduling
• Calendar Integration: Direct calendar service integration (Google, Outlook)
• Performance Optimization: Enhanced streaming assembly for large tool calls
• RAG: validation of task decomposition and estimation against industry relevant literature
• More granular task dependency: representation of tasks in a DAG instead of a list to allow overlap within projects, where feasible/convenient
• Input from GitHub issues: instead of processing markdown directly, it creates a list by parsing issue
• Reinforcement learning: training the agent to improve task decomposition and estimation from GitHub history (e.g. diffs in timestamps, issue comments etc.)

Prerequisites (Local/GitHub)

• Python 3.10
• Java 17+
• Docker (optional, for containerized deployment)
• Nebius API credentials (for LLM-powered features)

Installation

1. Clone the repository:
   hljs language-bash

   Copy

   git clone https://github.com/blackopsrepl/yuga-planner.git
   cd yuga-planner
   

2. Create a virtual environment:
   hljs language-bash

   Copy

   make venv
   

3. Install dependencies:
   hljs language-bash

   Copy

   make install
   

4. Set up environment variables / secrets:
   hljs language-bash

   Copy

   make setup-secrets
   # Then edit tests/secrets/cred.py to add your API credentials
   

5. Run the app:
   hljs language-bash

   Copy

   make run
   

Docker (Local/GitHub)

1. Build the image:
   hljs language-bash

   Copy

   docker build -t yuga-planner .
   

2. Run the container:
   hljs language-bash

   Copy

   docker run -p 7860:7860 yuga-planner
   

Python Dependencies

See requirements.txt for full list.

License

This project is licensed under the Apache 2.0 License. See LICENSE.txt for details.

Acknowledgements

• Hugging Face
• Gradio
• Nebius LLM
• llama-index
• Timefold