Operant

Operant is a computer use agent for applications that expose no API. A model drives the real screen through the macOS accessibility tree until the goal is reached, recording what it did. That recording is compiled into a capability with typed inputs and outputs, the path of steps, how each control is located, checkpoints, and per-tenant settings. Replay then runs with no model involved: it resumes from wherever the app already is, re-localizes the live screen by content, and path-finds straight to the goal when earlier steps are already done. Every risky, mutating, or sensitive step pauses for human approval. During discovery the model never sees a credential value; it requests one by name through a secret handle and a human supplies the value or names an environment or Keychain source, resolved late at replay time.

Technologies Used

Python

TypeScript

Node.js