Notte

Notte takes a different approach to browser automation for agents — instead of screenshots and vision models, it converts web pages to structured, text-based representations that agents can reason about directly (similar to accessibility trees but optimized for LLM consumption). This reduces cost (no vision API calls) and improves reliability for structured pages. Python-first SDK with clean action primitives for navigation, clicking, and form filling. Confidence is docs-derived.

Python SDK: pip install notte-sdk. Core pattern: env = notte.Session(); obs = env.step("goto https://example.com") returns structured page observation (interactive elements, text content, current URL). env.step("click Login") or env.step("fill email=user@example.com") for actions. Structured observations include element lists with labels, types, and identifiers — no CSS selectors or XPath needed. Supports session reuse for multi-step workflows.

API key auth for Notte Cloud. Keys from Notte platform. HTTPS enforced. Local mode: no API key, uses local Chrome. Session-level isolation between agent workflows.

Structured page representations may miss dynamically rendered content (SPAs with heavy client-side JavaScript). Fallback: env.screenshot() available for vision-model-based decisions when structured approach is insufficient. SDK failures surface as Python exceptions. Cloud uptime tracked at notte status page.

docs.notte.cc covers SDK quickstart, session management, action primitives, and integration with LangChain/LlamaIndex. Getting started: pip install notte-sdk, first structured page observation in minutes. Community via Notte Discord.