Ocular

Autonomous web agents typically rely on runtime visual-language model (VLM) loops: every step captures a screenshot, sends it to a remote model, and waits for coordinate predictions. This runtime dependency introduces latency, ongoing API expenses, and non-deterministic behavior on standard workflows. Ocular shifts model inference entirely out of the execution loop into an offline generation pass. Human demonstration traces are analyzed once by Meta models to synthesize an explicit state machine and extract robust visual anchors. Ocular then outputs a standalone Playwright driver wrapped in an authenticated FastAPI microservice. Key Characteristics: • Zero Runtime Inference: Executes as native Playwright automation with no active LLM or VLM network calls during execution. • Native Playwright Performance: Actions dispatch deterministically through local browser automation rather than remote multimodal loops. • Canvas & Non-DOM Support: Automatically falls back to scale-invariant geometric centroids when DOM selectors are unavailable or obscured within HTML5 canvas elements. • Execution Safety: Sandboxed through an in-process egress proxy (NetGuard) with automated sensitive parameter redaction, verified across a 145-test suite.