---
title: BirdScope AI - MCP Multi-Agent System
emoji: 🦅
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 6.0.1
python_version: 3.11
app_file: app.py
pinned: false
license: mit
short_description: AI-powered bird identification with MCP multi-agent system
tags:
  - building-mcp-track-enterprise
  - building-mcp-track-consumer
  - building-mcp-track-creative
  - mcp-in-action-track-enterprise
  - mcp-in-action-track-consumer
  - mcp-in-action-track-creative
---

# 🦅 BirdScope AI - MCP Multi-Agent System

**AI-powered bird identification with specialized MCP agents**

Built for the [MCP 1st Birthday Hackathon](https://huggingface.co/MCP-1st-Birthday)

---

## 📢 Hackathon Submission

**Social Media:** [Twitter/X Post](https://x.com/zulucoconuts/status/1995255281064755708)

**Demo Video:** [Watch on YouTube/Loom](https://youtu.be/V_ZoOkyjEyU)

**Track Submissions:**
- 🔧 **Track 1 (Building MCP)**: Two custom MCP servers
  - **Nuthatch MCP Server** - 7 tools for bird species database (search, species info, images, audio, family search, conservation filtering)
  - **Modal Bird Classifier MCP** - 2 Modal-hosted GPU-powered image classification tools (base64 & URL inputs)
  - Categories: Enterprise (wildlife conservation) | Consumer (bird enthusiasts and education) | Creative (multimedia exploration)
- 🤖 **Track 2 (MCP in Action)**: Full multi-agent system with supervisor routing
  - LangGraph-based supervisor orchestrating 3 specialized subagents
  - Integrates both MCP servers with intelligent tool routing
  - Categories: Enterprise (conservation orgs) | Consumer (bird watchers) | Creative (educational multimedia)

**Author:** [@facemelter](https://huggingface.co/facemelter)

**Built with:** Gradio 6 | LangGraph | FastMCP | Modal (GPU) | OpenAI/Anthropic/HuggingFace LLMs

---

## 🌐 Project Overview

BirdScope AI showcases an advanced multi-agent system powered by **Gradio 6** and **LangGraph**, designed to identify bird species, explore multimedia content, and provide educational information about birds worldwide.

**Our innovation:** We built **two complete systems in one**:
- 🔧 **Two Custom MCP Servers** (Track 1): Nuthatch species database (7 tools) + Modal GPU classifier (2 tools)
- 🤖 **Multi-Agent Application** (Track 2): Supervisor-orchestrated specialist agents

This dual approach demonstrates both **building MCP infrastructure** and **leveraging MCP for autonomous agents**.

---

## ✨ Key Features

### 🤖 Multi-Agent Orchestration
- **LangGraph Supervisor Pattern** with intelligent LLM-based routing
- **3 Specialized Subagents** (Image Identifier, Species Explorer, Taxonomy Specialist)
- **Session-based Agent Caching** - Agents reused within user sessions for 10x faster responses
- **Provider-Specific Prompts** - Optimized system prompts for OpenAI, Anthropic, and HuggingFace

### 🔧 Dual MCP Server Architecture
- **Modal Bird Classifier** ([modal.com](https://modal.com))
  - [prithivMLmods/Bird-Species-Classifier-526](https://huggingface.co/prithivMLmods/Bird-Species-Classifier-526) from HuggingFace
  - 526 bird species classification on Modal T4 GPU
  - Serverless GPU deployment for on-demand classification
  - Streamable HTTP transport with base64 and URL input support
- **Nuthatch MCP Server** (Custom Built - Track 1)
  - FastMCP framework with 7 specialized tools
  - Integrates [Nuthatch API](https://nuthatch.lastelm.software) (1000+ species)
  - **Dual Transport Support**: STDIO (subprocess) for HF Spaces + HTTP for local debugging
  - Data sources: Nuthatch DB, Unsplash (images), xeno-canto (audio)

### 📡 Dual Streaming Output
- **Chat Response Stream** - Real-time markdown rendering with embedded media
- **Tool Execution Log Stream** - Parallel visibility into MCP tool calls (inputs/outputs)
- **Async Progress Indicators** - Immediate user feedback before processing begins

### 🎨 Structured Output Parsing
- **LlamaIndex Pydantic Models** - Type-safe response formatting
- **Regex URL Extraction** - Automatic detection of image and audio URLs
- **Smart Audio Normalization** - xeno-canto links converted to browser-friendly format (`/download` → playable)
- **Markdown Media Embedding** - Images and audio automatically formatted

### 🌐 Multi-Provider LLM Support
- **OpenAI** (GPT-4o-mini) - Recommended for reliability
- **Anthropic** (Claude Sonnet 4) - Best for complex reasoning
- **HuggingFace Inference API** - Open-source models (limited tool calling)
- **User-Provided Keys** - No backend API key required, users supply their own

### 💅 Production UI/UX
- **Gradio 6.0 SSR** - Server-side rendering for enhanced performance
- **Custom Cloud Theme** - Sky-inspired CSS with mobile-responsive design
- **Dynamic Examples** - Example queries adapt to selected agent mode
- **Instant Feedback** - "⏳ Starting..." indicator appears immediately on submit

---

## 🗂️ Data Sources & MCP Servers

We built **two custom MCP servers** that integrate with bird data APIs and GPU-powered classification:

**Data Sources:**
- **Nuthatch API** ([nuthatch.lastelm.software](https://nuthatch.lastelm.software)) - 1000+ bird species database by Last Elm Software
- **Unsplash** - High-quality reference images for visual identification
- **xeno-canto.org** - Community-contributed bird audio recordings worldwide
- **HuggingFace Model** - [prithivMLmods/Bird-Species-Classifier-526](https://huggingface.co/prithivMLmods/Bird-Species-Classifier-526) for GPU classification

**MCP Servers:**
1. **Nuthatch MCP Server** (Track 1 - Building MCP)
   - 7 specialized tools: search, species info, images, audio, family search, conservation filtering
   - STDIO transport for HF Spaces, HTTP option for local debugging
   - FastMCP framework with async API integration

2. **Modal Bird Classifier** (GPU-powered)
   - Image classification tools: URL and base64 input support
   - Serverless GPU deployment via Modal
   - Streamable HTTP transport

---

## 🧩 Core Components

**Multi-Agent Orchestration:**
- **LangGraph Supervisor Pattern** - LLM-based routing between specialist agents
- **3 Specialized Subagents** - Each with focused tool subset (image ID, species exploration, taxonomy)
- **Session-based Caching** - Agent instances reused within user sessions for performance
- **Dual Streaming** - Parallel chat response + tool execution log streams

**Agent Architecture:**
- `subagent_supervisor.py` - Creates supervisor workflow with LangGraph
- `subagent_factory.py` - Builds specialists with filtered tool access
- `subagent_config.py` - Defines agent modes and tool allocations
- `prompts.py` - Provider-specific system prompts (OpenAI, Anthropic, HuggingFace)

**UI & UX:**
- **Gradio 6.0** with SSR for enhanced performance
- Custom cloud-themed CSS with mobile-responsive design
- Dynamic examples that adapt to agent mode selection
- Immediate processing feedback with async streaming updates

---

## 🚀 Quick Start

**Try the Live Demo:** Just provide your LLM API key (OpenAI, Anthropic, or HuggingFace) in the sidebar and start exploring!

**For Developers:**
```bash
# Clone and install
git clone <repo-url>
cd hackathon_draft
pip install -r requirements.txt

# Configure environment
cp .env.example .env
# Edit .env with your API keys

# Run locally
python app.py
```

**Deploy to HuggingFace Spaces:**
```bash
python upload_to_space.py
# Configure Secrets in Space Settings (see docs/dev/main-README.md)
```

**Full Setup Guide:** See [docs/dev/main-README.md](docs/dev/main-README.md) for comprehensive deployment instructions

---

## 🏆 Credits & License

Built for the [HuggingFace MCP 1st Birthday Hackathon](https://huggingface.co/MCP-1st-Birthday)

**Data Sources:** [Nuthatch API](https://nuthatch.lastelm.software) (Last Elm Software) | [xeno-canto.org](https://xeno-canto.org) | [Unsplash](https://unsplash.com)

**Technology:** [Model Context Protocol](https://github.com/anthropics/mcp) | [LangGraph](https://github.com/langchain-ai/langgraph) | [Gradio 6](https://gradio.app) | [Modal](https://modal.com)

MIT License - Educational and research purposes