Kevinweisl/browser-agent

A generic browser automation agent that turns natural-language task descriptions into real Chromium actions using Playwright plus Claude/NIM LLMs. Intended for developers/researchers who need robust, LLM-guided web workflows (origin: AI Coding Test 2026 interview deliverable). Key features: - Planner → Actor → Validator loop: LLM planner issues steps, Playwright executes, LLM validator approves; failures trigger replanning. - Self-correction: K=2 validator voting rejects false positives and forces replan instead of blind retries. - Self-maintenance: 6-tier locator ladder (cached → role → label → testid → text → CSS/no-class → vision-stub), caching in Postgres, and Healenium-style aria fingerprints to heal selector drift. - Eval: 10/10 on a 10-task suite (SEC, Reuters, finance pages). - Includes FastAPI demo UI and Zeabur deployment instructions. Limits: single-task concurrency, in-memory registry, 5–300s per-task caps. License: MIT.

AI-related keywords found in this repo's description, topics, or README summary — grouped by category. Each badge links to the corresponding ranking detail page.