Pietro Schirano is founder of MagicPath and a designer and AI-tool developer. His work connects interactive prototyping with tools that generate software and let people refine the result through a visual interface.[1][2]
Start with Claude Engineer v3 →
Career
In his Dive Club interview, Schirano traces product-design work at OpenTable through Facebook, Uber and Brex. Facebook’s prototyping culture taught him to put an interaction into someone’s hands before deciding it worked. At Brex, early GPT-3 access led him to prototype an AI financial interface, including graphs generated through SQL output. That experience informed his subsequent work on AI creation tools. He also discusses design work with Anthropic.[3]
Contributions
Create and load tools during a conversation
Claude Engineer v3 combines a command-line or web interface with dynamic Python tool creation. A tool-creator component generates a module for a requested capability, and the framework loads and invokes it during subsequent work. Conversation history and token tracking provide the surrounding context. Its documented setup uses Anthropic credentials, with E2B access for sandbox features. The repository exposes how generated capabilities enter the running assistant, making the execution path available for inspection.[4]
Keep generated interfaces open to direct manipulation
Schirano’s current MagicPath descriptions present a shared canvas for code, interactive components, diagrams and flows. An agent can bring repository components into that canvas, while a visual editor supports direct resizing, duplication and other changes. His Figma-import account describes converting frames into React code with layout information retained. The underlying idea is to connect agent generation with an interface where a designer can keep adjusting the working result.[2]
Tools, writing & talks
Tools
Start here
Claude Engineer v3
github.comA Python assistant framework that generates and dynamically loads tools, with command-line and web interfaces.
Read how new tools are created, loaded and called, then inspect their code before execution. Python familiarity and Anthropic access are needed; sandbox features also use E2B. The repository documents a Claude 3.5 configuration.
Creator and repository author[4]Original source