Om sesjonen
Most AI agent tutorials stop at "call the API and print the response." That's not an agent — that's autocomplete with better marketing.
In this workshop you build a real agent system: three services in Docker that work together so the AI can discover and call external tools on its own. The protocol that makes this work is MCP (Model Context Protocol), the open standard from Anthropic that's already integrated into Claude Desktop, VS Code, and Cursor.
You'll build a chat UI, an orchestration layer (Python/FastAPI + OpenAI), and an MCP server where you register your own tools. The agent discovers what's available at startup via JSON-RPC — add a tool, restart, it picks it up. No glue code.
After a short architecture intro, you code. Environment runs in GitHub Codespaces, so setup is a fork and a docker compose up. You'll build two tools from scratch, debug failures, and watch the agent combine multiple tools in one response.
The workshop uses Python, but the MCP protocol and JSON-RPC patterns are language-agnostic — they map directly to any JVM language. For developers and architects who want to understand how AI tool-calling works under the hood, not just consume an API.
Bring a GitHub account and basic programming skills. You leave with a running system to fork and extend.