FreshEyes

You can't un-see your own product. FreshEyes can — an AI agent walks your site like a first-time visitor and shows you exactly where they get stuck.

Comment

FreshEyes

An AI agent that walks your website like a brand-new visitor and reports exactly where they get stuck — with concrete fixes and screenshot evidence.

🔗 Try it

Mission

Everyone ships now — but shipping isn't the finish line; being understood is. FreshEyes exists to give every builder the one thing they can't get on their own: an honest, first-time look at their own product. We want a stranger's confusion to be visible before it costs you a user — so anyone, on any budget, can see their product through fresh eyes and fix the first impression early.

Inspiration

Everyone ships now — developers, PMs, and designers can all build and put a site live. But you can't un-see your own product: the people who build a website already know what it is, who it's for, and where every button leads. So they're the worst judges of the first 10 seconds a stranger spends on it, even though those 10 seconds decide whether the stranger stays.

The usual ways to close that gap are bad: real user testing is slow and expensive — and by the time that feedback reaches you, you've already lost users; "audit" tools just run Lighthouse and hand you performance scores; and asking friends gets you politeness instead of the moment they got lost. There's a gap between "technically works" and "a real person lands on it and actually gets it." FreshEyes lives in that gap — it sends an AI through your site as a confused newcomer and tells you what it found.

What it does

Paste any URL. FreshEyes:

  • Opens a real cloud browser and visits your site as a first-time visitor — no logins, no assumptions.
  • Streams every action live — you watch it look, click, and react in real time, with its reasoning shown as it goes.
  • Judges the whole first impression across 11 dimensions — clarity, CTA, visual design, imagery, copy, navigation, trust, forms, accessibility, performance, and errors.
  • Records each friction point with a severity, the page URL, a screenshot as evidence, and a concrete fix (actual colors, button styles, copy — not "improve the design").
  • Returns a Markdown report you can read inline, scrub frame-by-frame, and export to Markdown or PDF.

Crucially, it's calibrated, not padded: it first figures out what kind of page this is — a throwaway placeholder vs. a real product page — and matches the depth and severity of its audit to reality.

How I built it

FreshEyes is a brain + hands agent. An LLM does the reasoning; a real cloud browser does the doing; everything streams to the UI live.

  • Brain: the OpenAI SDK pointed at OpenRouter, running a hand-rolled, bounded tool-calling loop. The model decides the next browser action one step at a time.
  • Hands: Stagehand on Browserbase — a managed remote browser with a live view. The model's tools are observe, act, screenshot, record_finding, and finish.
  • The rubric is a file: the agent's judgment lives in skill.md, a first-time-visitor audit guide loaded into the system prompt — so the product's "taste" can be tuned without touching code.
  • Backend: Bun + Express, streaming the agent's steps, reasoning, screenshots, and findings to the browser over Server-Sent Events.
  • Frontend: Next.js + Tailwind — a live browser view, a streaming activity feed, and a Markdown report with PDF/Markdown export.
  • Analytics — Novus: the app is instrumented with Novus (Pendo's product agent, the hackathon's sponsor tool). Novus connected to the repo, auto-detected pages and click events, and instruments the full audit funnel (audit_started → audit_completed) plus AI-agent analytics on the audit interaction itself — so real user behavior is measurable the moment the first stranger lands, without hand-writing tracking code. Novus even raised a signal on the gap between audits started and completed — which pointed straight at the session-timeout report-delivery bug I then fixed, so the analytics directly drove a product improvement. Novus dashboard proof: https://github.com/rushibhosalepro/fresheyes/tree/main/novus-proofs

Challenges I ran into

  • Driving a real browser reliably. An agent clicking through arbitrary sites fails in a hundred ways — logins, bot-checks, timeouts. The fix was to treat a block as a finding (not a crash), cap the loop, capture the live view immediately so it never looks stuck, and guard the run so a dropped connection can never re-trigger a second audit.
  • Surviving the session timeout. A long audit can hit the cloud browser's session cap mid-run. Instead of failing, FreshEyes builds the final report from everything gathered so far and reliably delivers it to the UI even after the live stream drops.
  • Making the model behave reliably as an agent. Models can emit malformed tool calls. The loop had to feed every error back as a message the model could read and self-correct from, instead of dead-ending.
  • Proportionate judgment. Early versions over-reported — four findings and a "HIGH" on a placeholder page. Moving the calibration logic into skill.md ("figure out the page type first, then audit to that") was the biggest jump in quality.
  • Streaming an agentic loop to the UI. Surfacing the agent's reasoning, actions, screenshots, and findings as distinct live events — and letting the user truly cancel mid-run — took real iteration.

What I learned

  • The rubric is the product. A plain "review this site" prompt produces generic, padded feedback. A calibrated skill.md is what turned it into something useful.
  • A defensive loop matters more than the model. Feeding tool errors back so malformed calls self-correct is what made the agent reliable, and routing through OpenRouter let me swap models with a one-line change.
  • Reliability beats cleverness for a demo. Live view, block-as-finding, a real Stop, and a run-once guard are what keep a stranger's URL from ever hitting a blank screen.
  • Watching is the magic. Streaming the agent's actions live turned "an AI looked at your site" into "I watched an AI get confused by my site" — far more convincing.

What's next

  • Background jobs for long audits — large sites can take anywhere from ~5 minutes to an hour to crawl fully. Move audits onto a durable background job queue with persistent run storage, so a scan keeps running independent of the browser connection and users can close the tab and come back to a finished report.
  • Vision by default — pair a vision-capable model with the existing screenshot path so visual-design and imagery findings come from real pixels, not the DOM.
  • Multi-page journeys — follow a full funnel (landing → pricing → signup), not just the first screen.
  • Before/after re-audits — track a score over time as you ship fixes.
  • Persona modes — audit as a "skeptical buyer," a "mobile user on slow data," or an "accessibility-first" visitor.
  • Shareable report links — a public URL per audit to hand to your team.

Built With

  • browserbase
  • bun
  • events
  • express.js
  • jspdf
  • next.js
  • novus
  • openai
  • openrouter
  • pendo
  • react
  • react-markdown
  • sdk
  • server-sent
  • stagehand
  • tailwind
  • vercel
Share this project:

Updates