github.comOpen source

fresheyes

An AI agent that uses your website like a confused first-time visitor — narrating every point of friction, scoring the experience, and proposing fixes.

fresheyestool.comOpen Project
fresheyestool.comView Repository
Project Overview

fresheyes is an open-source CLI tool that points a vision-capable AI agent at any web app and has it attempt a real user goal ("understand what this does and sign up") the way a skeptical newcomer would — by sight, not by reading the code. As it navigates, it narrates its confusion and decisions in real time, then produces a self-contained narrated HTML replay of the session plus a "rage-quit score" anchored to pre-generated, binary-checkable milestones. Beyond diagnosis, it generates tiered fixes — copy, layout, even SVGs — and in repo-aware mode proposes a reviewable change plan as a diff, never auto-editing. It supports both pre-login and logged-in reviews via locally saved browser sessions (credentials never touch the agent), and is extensible through markdown personas and MCP integrations like Figma and shadcn. The capstone: fresheyes built its own marketing site, then reviewed that site using itself.

Tech Stack
TypeScriptNode.jspnpm workspacePlaywrightClaude (vision agent)MCPCLI
Key Features
  • Vision-first agent that operates a live web app like a real first-time user, using a hybrid vision + element-map acting model
  • Real-time narration of friction, confusion, and decision-making as it navigates
  • Self-contained narrated HTML replay of every run — no video encoding or ffmpeg needed
  • Rage-quit score anchored to 5 binary-checkable milestones generated before the run, with a --reuse-milestones flag for fair before/after comparisons
  • Tiered fix generation: copy rewrites, layout suggestions, and generated SVGs
  • Repo-aware change plans — proposes a reviewable diff, never auto-edits your code
  • Pre-login and logged-in modes, with sessions saved locally and credentials kept out of prompts entirely
  • Swappable markdown personas (~20 lines each) as the main contribution surface
  • MCP extensibility for design tokens, UI components, and media generation
  • Built-in safety caps (--max-steps, --max-cost) and a consent gate

© 2025 Werner's Works. All rights reserved.