Expand description
LLM-as-DOM: AI browser pilot with cheap LLM + heuristics.
A headless browser pilot that compresses web pages to ~100-300 tokens and uses heuristics + a cheap LLM to accomplish goals autonomously.
Re-exports§
pub use error::Error;
Modules§
- a11y
- Accessibility tree extraction via JS injection.
- audit
- Page quality audit engine.
- backend
- LLM backend implementations for the browser pilot.
- cloaking
- CSS cloaking heuristics + SPA shell detection.
- crypto
- Chrome cookie decryption (macOS Keychain + AES-128-CBC).
- engine
- Browser engine abstraction layer.
- error
- heuristics
- Rule-based action engine – resolves 70-90% of actions without LLM.
- intent
- SS-3: Centralized goal intent parser.
- locate
- Source location engine for mapping DOM elements to source files.
- network
- Network traffic capture and semantic classification.
- oauth
- OAuth flow detection, provider identification, and flow state tracking.
- observer
- Observer module for DOM diffing and monitoring.
- pilot
- Browser pilot: observe -> heuristics -> LLM fallback -> act loop.
- playbook
- Playbook system: deterministic step-by-step replay for known workflows.
- profile
- Chrome profile discovery and cookie loading.
- sanitize
- Steganographic prompt-injection defense.
- selector
- Semantic selector: find elements by description, not CSS.
- semantic
SemanticView: compressed DOM representation for LLM consumption.- session
- Session state: cookies, navigation history, and page memory across navigations.
- target
- Semantic element targeting.
- watch
- Watch system: persistent page monitoring with semantic diffing.