Jiri 🤖

Self-Improving AI Agent with Dynamic Tool Discovery

An intelligent agent that learns and evolves in real-time by automatically discovering and adding new MCP (Model Context Protocol) tools as needed. Start with zero capabilities — Jiri builds its own toolkit on-demand by semantically searching for relevant servers, connecting them, and expanding its abilities with every query.

---

🌟 The Self-Improving Difference

Unlike traditional AI assistants with fixed capabilities, Jiri:

• Starts with zero tools — Lightweight and fast to initialize
• Discovers tools at runtime — Semantic search finds the right MCP server for any query
• Chains tools automatically — Seamlessly combines multiple tools for complex multi-step tasks
• Builds its own toolkit — Automatically connects to new capabilities as users ask questions
• Remembers what works — LRU cache keeps frequently-used tools loaded
• Gets smarter over time — Usage metrics learn which tools to preload on next startup
• Handles failures gracefully — Unhealthy servers get cooldowns, working servers persist

---

📁 Two Implementations

This repository contains two separate implementations of Jiri's MCP architecture, side-by-side:

	Dedalus	LangChain
Agent Framework	Dedalus Labs SDK	LangGraph + langchain-mcp-adapters
LLM Provider	Anthropic (Claude Haiku 4.5) via Dedalus	OpenAI (GPT-4.1-mini) via LangChain
Embeddings	Dedalus Embeddings API	OpenAI text-embedding-3-small
MCP Transport	Dedalus marketplace URLs	Direct HTTP/SSE/stdio connections
Tool Discovery	Dedalus marketplace semantic search	Local semantic search with OpenAI embeddings
Custom MCP Servers	—	✅ (e.g. servers/news_server.py via stdio)
API Key Required	DEDALUS_API_KEY	OPENAI_API_KEY

Both implementations share the same core architecture:

MCP/
├── web_server.py           # FastAPI web UI with WebSocket
├── static/index.html       # Chat interface
└── router/
    ├── core.py             # SmartRouter orchestrator
    ├── registry.py         # Tool registry with semantic search
    ├── tool_cache.py       # LRU cache for active servers
    ├── health.py           # Server health tracking
    ├── metrics.py          # Usage analytics
    ├── history.py          # Conversation history
    └── config.py           # Configuration

---

🚀 Quick Start

Dedalus Implementation

cd Dedalus
cp .env.example .env
# Add DEDALUS_API_KEY to .env
uv sync
cd MCP && uv run python web_server.py

LangChain Implementation

cd LangChain
cp .env.example .env
# Add OPENAI_API_KEY to .env
uv sync
cd MCP && uv run python web_server.py

Then open: http://localhost:8080

---

🎯 How It Works

Runtime Tool Discovery

Jiri doesn't come pre-configured with tools. It discovers capabilities dynamically based on what you need:

You: How is MSFT stock doing?

1. Router checks cache → No stock server found
2. Semantic search discovers a finance MCP server
3. Connects and executes stock lookup
4. Returns real-time MSFT data ✅

Cache now contains: [finance-mcp]

The Self-Improving Cycle

First Session (Cold Start):
  Query 1: "MSFT stock"      → Discovers finance tool → 3s
  Query 2: "AAPL stock"      → Uses cached tool       → 1s ⚡
  Query 3: "Send email"      → Discovers email tool   → 3s

Second Session (Learned Preferences):
  Startup: Preloads finance (most used) ← AUTOMATIC!
  Query 1: "TSLA stock"      → Uses preloaded tool    → 1s ⚡

Automatic Tool Chaining

You: Explain the TensorFlow GitHub repo and email the summary to my team

1. Discovers Deep Wiki tool → Analyzes repository
2. Discovers Gmail tool → Sends summary email
✅ Multi-tool workflow from a single natural language request!

---

✨ Core Features

• 🧠 Runtime Tool Discovery — Zero configuration, discovers MCP servers on-demand
• 🔗 Automatic Tool Chaining — Chains multiple tools for complex multi-step tasks
• 🔄 Continuous Learning — Usage patterns shape which tools get preloaded
• 🗄️ LRU Caching — Keeps frequently-used tools loaded for instant reuse
• ❤️‍🩹 Adaptive Health — Failed servers get cooldowns, system self-heals
• 🎨 Beautiful Web UI — Real-time chat with live logging and cache visualization
• 📊 Live Observability — Watch tool discovery, cache updates, and execution in real-time

---

🏗️ Architecture

User Query
    ↓
SmartRouter.handle_turn()
    ↓
Check cache for matching tools
    ↓
If not found → discover_tools() → Semantic search → Add to cache
    ↓
Execute tool via MCP server
    ↓
Post-run: LRU touch, metrics, health tracking
    ↓
Return response to user

Components:

Component	Description
SmartRouter	Main orchestrator — manages discovery, caching, and execution
ToolRegistry	Semantic search over MCP registry using embeddings
ToolCache	LRU cache for active MCP server connections
HealthTracker	Server failure tracking with automatic cooldowns
UsageMetrics	Persistent usage analytics for smart preloading
ConversationHistory	Multi-turn dialogue context with rollback support

---

📖 More Details

• Dedalus README — Setup, configuration, and usage for the Dedalus Labs implementation
• LangChain README — Setup, configuration, and usage for the LangChain/LangGraph implementation

---

🔒 Security

• API keys stored in .env (gitignored)
• No credentials in logs or agent context
• WebSocket connections are local only

---

📄 License

MIT

---

Built with ❤️ for TartanHacks 2026