A step-by-step guide to building a privacy-preserving AI agent that orchestrates medical image de-identification without ever exposing patient data to the model. Using FastMCP and the MCP Python SDK, you expose six tools from the Aegis DICOM de-identification pipeline (built on MONAI/PyTorch) to Claude Desktop. The agent can warm up models, process single files or entire directories in background threads, poll job status, and summarize audit reports — all through natural language conversation. Key design principles include returning only summary statistics (never pixels or PHI text), using async background jobs for long-running tasks, and maintaining a verifiable on-disk audit trail. The tutorial also covers testing with MCP Inspector, configuring Claude Desktop via JSON, and security considerations like restricting file access and avoiding PHI in file names.
