In Claude Code, the plugin runs all of these on your subscription ($0), as plain-English asks: "run fresheyes on X and propose fixes", "run fresheyes on X and plan the fixes against my repo at ../your-app", "review X with the senior-product-designer lens". The hub is a standalone browser UI. The npx forms below are the standalone path with your own key (or the offline mock brain for the bundled demo).

# propose fixes for the ranked frictions (annotated screenshots plus a rebuilt screen)
npx fresheyes https://your-app.com --goal "sign up" --fix

# map frictions into your codebase as PR-style diffs (read-only, never applied)
npx fresheyes diff ./fresheyes-runs/<run> --repo ../your-app

# the expert design pass: every AI-design tell named, plus a full-page rebuild
npx fresheyes critique https://your-app.com --lens senior-product-designer

# everything in a browser: launcher, live narration, reports, personas, connections
npx fresheyes hub

# review the logged-in product: YOU sign in, it keeps only the cookies
npx fresheyes login https://your-app.com
npx fresheyes https://your-app.com --goal "..." --session

# no key needed: run the whole pipeline offline against the bundled demo app
npx fresheyes http://localhost:4173 --goal "sign up" --brain mock --fix

1. Sees your app by sight (a real screenshot, like a human, not your DOM or code).
2. Tries to accomplish the goal as a specific persona, narrating its honest inner monologue.
3. Flags friction: every hesitation, wrong click, dead end, and rage-quit moment.
4. Scores the run against five milestones fixed before it starts, so the number is reproducible (pin them and re-run for honest before/after numbers).
5. Fixes it: generates the better version, then reviews its own fix before showing you anything (does it actually remove the friction, match your conventions, and not break things). Anything that fails the self-review is withdrawn.
6. (Optional) plans the change against your real repo: reads your actual code and hands you a PR-style diff. Read-only by construction.

Before the run, the goal is decomposed into exactly five ordered, screen-checkable milestones (for example "found the path toward signup" up to "the goal is fully accomplished"). A milestone only counts when the screenshot proves it. The score is the count of milestones reached, 0 to 5. Because the milestones are fixed up front (and can be pinned and reused with --reuse-milestones), the score is comparable across runs, so you can measure whether a change actually helped.

• Pre-login: starts cold and reviews the public path (landing, then the way to signup). This is where strangers bounce.
• Logged-in (you bring the session): you sign in yourself in a headed browser, fresheyes keeps only the resulting session cookies, and it reviews the real product like a confused new user who is already through the door.

On login, fresheyes never automates credentials, Google sign-in, SSO, 2FA, CAPTCHAs, or payments. Those are deliberately bot-hostile and no open-source tool should drive your real accounts. You log in; fresheyes explores. (And when an app throws a sign-in wall at a stranger in the first ten seconds, the agent bouncing there is itself a finding.)

A persona is about twenty lines of markdown (no code) describing a real kind of user: how they behave, what annoys them, when they would quit. The bundled default is skeptical-newcomer (never seen your product, mildly distrustful, low patience for jargon); name it in your ask ("as the skeptical-newcomer persona") or add your own, and every install sees more. A couple of frontmatter knobs the runner actually honors (viewport: 390x844, patience: 0.5).

---
id: impatient-mobile-user
name: The Impatient Mobile User
one_line: On a phone, one-handed, distracted, will not zoom in to read.
viewport: 390x844
patience: 0.5
---

You are a real person visiting this app for the very first time...

Drop one in personas/ and PR it. The library grows every time someone adds a kind of user, and every install sees more.

Personas tell you where a stranger gives up. Lenses tell you what a senior professional would change. Same format, same contribution surface: markdown, no code. The lens runs on your subscription in Claude Code too ("review X with the senior-product-designer lens"); the npx critique form below is the standalone path with your own key.

npx fresheyes critique https://your-app.com \
  --lens senior-product-designer --lens conversion-specialist

Built-ins: senior-product-designer (hierarchy, type, spacing, craft), conversion-specialist (clarity, CTAs, trust, friction), and brand-designer ("with the logo removed, would anyone know this site is yours?"). Before critiquing, every lens loads a curated design knowledge base (measured patterns from genuinely great sites, design-system principles, conversion heuristics, and the named tells of AI-generated design). You get a craft score, an ai_verdict (clean / mixed / ai-made), ranked issues, and one full-page rebuild shipped as a self-contained before/after.

fresheyes works out of the box. Connections strengthen the fix step: fresheyes finds the friction and decides what to improve; connected tools give it better ways to produce the improvement.

• Higgsfield MCP for stronger hero art and image assets.
• Figma MCP to pull your real design tokens and components so fixes match your system.
• shadcn MCP to build fixes from clean, consistent components.
• Your repo plus a key for repo-aware change plans.

Secrets live in ~/.fresheyes/connections.json (chmod 600) and only ever leave as masked previews. Connected tools can propose during the fix step; they never write to your repo or app. See docs/EXTENSIONS.md.

No. fresheyes runs on Claude Fable 5 by default. On sites with biology, security, or similar sensitive-topic content, Claude's safety system may briefly switch to Opus 4.8 for part of the run (a yellow notice in Claude Code). This is expected, not an error; the review still completes. fresheyes does not and cannot override this, by design. See Cleaner prompts, fewer false interruptions for keeping run prompts clear, which reduces false trips without disabling anything.

Not for the plugin. Installed in Claude Code, runs are free on your Claude subscription ($0 per run). You only need your own ANTHROPIC_API_KEY for standalone or CI runs with --brain api. The offline mock brain (--brain mock) runs the entire pipeline with no key and no network.

No. Everything is a proposal. The browser run never writes anywhere. Repo mode is read-only by construction (the planner's only tools are readers) and hands you a diff you apply yourself, or do not.