Challenge — AI Assistant with Tool Use

Objective

Build a full-stack AI assistant that uses Claude with tool use to answer questions about real-world data — weather, database queries, or external APIs.

Requirements

Backend (50 points)

• Express or NestJS API with a POST /api/agent endpoint
• At least 3 tools defined and implemented:
  • Tool 1: Search/query your own database (e.g., products, users, posts)
  • Tool 2: Weather or external API call
  • Tool 3: A computational tool (e.g., calculate, format, convert)
• Proper agentic loop — handles multi-step tool use
• Parallel tool execution with Promise.all
• Error handling per tool — returns error message to Claude instead of crashing
• All tool inputs validated with Zod before execution

Frontend (30 points)

• Chat UI (React) — user can send messages and see responses
• Tool call visualization — show which tools were called and their results
• Streaming responses using SSE (from Module 23 streaming lesson)
• Message history preserved during session
• Loading state while agent is working

Quality (20 points)

• TypeScript throughout — tools have typed inputs/outputs
• Conversation history maintained correctly across multiple turns
• Rate limiting on the agent endpoint
• Error messages shown in the UI gracefully

Tool Ideas

Choose tools that make sense together:

Option A: Product Assistant

1. search_products(query, category?, maxPrice?) — Prisma
2. get_product_details(productId) — Prisma
3. check_stock(productIds[]) — mock or real
4. calculate_discount(price, couponCode) — computation

Option B: Travel Assistant

1. get_weather(city, days?) — OpenWeatherMap API
2. search_flights(from, to, date) — Amadeus or mock
3. convert_currency(amount, from, to) — ExchangeRate API
4. translate_text(text, targetLanguage) — LibreTranslate

Option C: Code Assistant

1. run_code(language, code) — execute in a sandbox
2. search_npm(query) — npm registry API
3. get_github_repo(owner, repo) — GitHub API
4. check_cve(packageName, version) — Snyk API

Grading

Criteria	Points
Agent loop correctly handles tool calls	15
3 tools defined with proper schemas	15
Tools actually execute and return data	15
Parallel tool execution	5
Chat UI functional	15
Tool call visualization	10
Streaming responses	5
Error handling (per tool + global)	10
TypeScript quality	5
Rate limiting	5
Total	100

Example Interaction

User: "I'm planning a trip from NYC to London next month.
       What's the weather like there, and how much is $2000 in GBP?"

Claude: [calls get_weather("London") and convert_currency(2000, "USD", "GBP") in parallel]

Claude: "London in [month] is typically cool and rainy — expect around 12°C with overcast
         skies. Your $2,000 budget converts to approximately £1,580 GBP at today's rate
         of 0.79. That should cover a comfortable week-long stay!"

Bonus

• Persist conversation history per user in the database
• Memory tool: remember(fact) and recall(topic) — stores and retrieves user preferences
• Multi-agent: one orchestrator Claude calls specialized sub-agents
• Deploy the agent to production (Railway + Vercel)