Grade page sections for clarity, writing and on-page SEO. WXT + TypeSafe AI Jev.
Made by Kitze
kitze.io · X · YouTube
 pwn.watch Live security incidents: CVEs, patch status, fixed versions and advisories. |
 flavors.design Copy-pasteable design flavors. Pick a vibe, hand the markdown to your agent. |
 gifs.so Search, copy, and download reaction GIFs. |
 Shipper Changelog, roadmap, feature requests and tasks for every product you ship. |
 Kanso One keyboard-first app for your todos, habits, routines, planner and saved stuff. |
 HealthMaxx One health log for everything you track: by hand, from wearables, or through your AI. |
 Zero To Shipped A full-stack starter kit for web and mobile apps. |
 Sotto Voice-to-text for macOS. Local AI, one-time purchase. |
 Tinkerer Club A private community for builders, self-hosters, and AI tinkerers. |
 Sizzy The browser for web developers. |
 Supermac A macOS command center for everyday workflows. |
 DMX A focused desktop client for X. |
 Matte 3D mockups and screen recordings with pro-level video editing |
 Postiz Agentic social scheduler for GIFs! |
 ScreenshotOne The best screenshot API for developers and agents |
 ether.fi a neobank with a built in ai-connector, build your own frontend to your finance with claude, chatgpt or your LLM of choice! Feedback always welcome! |
 OpenProse Reliable, event-driven multiagent workflows — in Markdown. |
 OpenTweet Agentic X (Twitter) autopilot. |
 AppSumo 1M+ entrepreneurs use AppSumo to find the best deals on software for their business. One-time payments, lifetime access. |
 Team Spartan by CONFLICT Know what matters. Get it fixed. Experience spans high-traffic consumer platforms, cloud infrastructure, and security — built under real operational pressure. |
 Browserless.io The web layer your agent runs on. |
 Strix Hack yourself before the bad guys do. Open source (70k+ stars on Github) |
 Maple VC We're 3 operators (digitalocean, uber, shopify) turned investors who write first checks into folks with ambition to build global businesses. |
 Rakazo Open source Grok Bot |
 The Influencer AI AI influencer generator: create a realistic influencer from traits or your photos, then make photos, talking videos and UGC ads with the same face every time. |
 Pluginboard Find the right plugins for your AI agents. Plugins, skills and step-by-step recipes for Claude, ChatGPT, Grokbot, Muse, Gemini and more. |
 file.cafe Free, instant file sharing. |
|
Your product here Reach developers in every kitze repo. |
||
A WXT Chrome extension that grades readable page sections with TypeSafe AI Jev through Vercel AI Gateway. Each section gets ten rubric scores; the page gets an A–E grade and a breakdown you can inspect.
MIT licensed. Bring your own Vercel AI Gateway key. No backend or analytics.
Requires Bun and Chrome 120+ (Manifest V3).
bun install --frozen-lockfile
bun run build- Open
chrome://extensionsand enable Developer mode. - Click Load unpacked and select
.output/chrome-mv3inside this checkout. - Open a normal web page and click PageGrade's toolbar icon. Chrome's native side panel opens alongside the page (its side follows your Chrome preference).
- Save your Vercel AI Gateway API key under Connection. Gateway credits and Jev access may be required.
- Click Analyze sections to consent to sending that page's extracted content for evaluation.
Section scores appear progressively in the side panel and as A–E badges on the actual page. Hover or focus a badge to outline its section; click it to select the breakdown in the panel. Show scores on page toggles overlays. Badges track scrolling/resizing; changed text invalidates affected grades.
The panel follows the active tab. Each tab keeps its own session report. Click the toolbar icon on a newly selected page to grant activeTab access if prompted. No all-sites access is requested.
Browse grades, expand rubrics, use Locate on page, or Export JSON. Stop retains completed section scores; an individual request failure does not discard other sections. Clear removes report and overlays. Source navigation or closing clears its report.
Assess whole page is optional, available after every extracted section is scored and the full text fits within 60,000 characters. Larger pages still get their section scores, a clearly labeled scored sections average, word/section counts and local on-page checks—not a fabricated whole-page grade.
bun run zip creates the unpacked archive under .output/. This is not a Chrome Web Store listing or a signed CRX.
Jev answers typed score questions with five explicit ordered anchors. Fractional scores are valid: normalize its 0–4 result to 0–100 by multiplying by 25. These values are rubric positions, not confidence probabilities.
| Metric | Weight |
|---|---|
| Clarity | 15% |
| Concision | 10% |
| Specificity | 10% |
| Explanation | 15% |
| Usefulness | 15% |
| Readability | 10% |
| Flow | 8% |
| Claim support | 7% |
| Writing quality | 5% |
| Search intent | 5% |
Exact questions, all five anchors, and fixed improvement guidance live in lib/rubric.ts. Jev does not generate rewrites or free-form advice.
sectionQuality = weighted mean of section scores
section weight = sqrt(min(max(wordCount, 1), 500))
composition = 30% section structure
+ 25% length fit
+ 20% focus / non-repetition
+ 25% purpose coverage
overall = round(70% sectionQuality + 20% composition + 10% onPageChecks)
Length and section count are judged against the page's apparent purpose. No minimum word count, no universal ideal section count, no reward for padding. A brief landing page and a long guide can both score well. Capped square-root weighting prevents both tiny fragments and one huge section from dominating.
Local on-page checks equally weight title presence, description presence, H1 convention, heading hierarchy, image alt-attribute presence and document-language presence. They do not claim to predict search rankings. Empty image alt is allowed for decorative images; meaningful alternative text is not evaluated.
| Grade | Score | Color |
|---|---|---|
| A | 85–100 | Dark green |
| B | 70–84 | Light green |
| C | 55–69 | Yellow |
| D | 40–54 | Orange |
| E | 0–39 | Red |
Editorial heuristic only. Not an official Nutri-Score, fact-check, comprehensive SEO audit, accessibility audit or ranking prediction. No affiliation with Nutri-Score. The rubric is not calibrated against human ratings yet. Scores may vary between runs.
- Reads the first
main/[role=main], otherwise the firstarticle, otherwisebody. - Uses headings and semantic section boundaries, without counting nested text twice.
- Excludes navigation, footers, sidebars, forms, editable fields, controls, hidden content, scripts and iframes.
- Does not inspect images/video meaning, canvas, shadow DOM or unloaded content.
- Long sections are split at word boundaries into 6,000-character chunks; no ordinary section text is dropped. Ten metrics are evaluated for every chunk, then averaged by character count. The panel discloses chunked sections: these are passage-level judgments, not an assessment of cross-chunk flow.
- Extraction safety guards: 250 sections, 2 million characters total, 100,000 visited DOM nodes. Hitting a guard is disclosed and prevents whole-page assessment. Coverage refers only to extracted text.
- At most two model requests concurrently across section workers; chunks within a section run sequentially. One active page analysis globally. No automatic retries, background scanning or automatic paid work on tab switches.
- Every required answer must be present, correctly typed, finite and in range. One failed chunk marks that section failed; other section results survive. Failed optional page assessment also preserves section scores. No fabricated fallback scores.
- Text changes invalidate affected displayed grades. A full snapshot check at completion rejects a changed page; navigation clears the report. A worker restart shows an interruption message with completed section scores retained.
- Click-to-analyze only. Opening the panel extracts locally; network evaluation starts only after clicking Analyze sections or Assess whole page.
- Sends extracted main-page text, section headings, page title, description and language to Vercel AI Gateway, which routes to TypeSafe AI Jev. Do not analyze sensitive pages. Provider retention policies apply; this project does not promise zero retention.
- Does not send the source URL, query string, DOM selectors, form values or browser history to the model. Text and metadata themselves can still contain sensitive information.
- API key stays in extension-local storage restricted to trusted extension contexts. It is not encrypted, synced, bundled or sent to source pages. Remove it under Connection.
- Scores, headings and source origin are held in browser session storage, cleared on source navigation/close, browser exit or Clear. Extracted source text is not persisted in reports.
- Exported JSON includes scores, headings and source origin; review before sharing.
- Permissions:
sidePanel,activeTab,scripting,storage, and access tohttps://ai-gateway.vercel.sh/*. No persistent all-sites permission, external messaging or remote code. - Page content is treated as untrusted evidence in all model questions. Model output never runs code or follows links. Isolated Shadow DOM overlays display validated scores without rewriting source content. Page-side messages can only select or invalidate known sections, never access keys or start paid analysis. Locate briefly highlights the selected section.
bun run check # TypeScript, lint, focused tests, formatting
bun run build
bun run zipRun the optional live API smoke with a key supplied through your shell environment:
bun scripts/smoke-jev.tsThe smoke sends only a hardcoded synthetic basil-growing guide. It requires AI_GATEWAY_API_KEY and prints metrics/counts, never the credential. SMOKE_REPORT_PATH optionally writes its synthetic report locally.
No GitHub Actions are configured. No production secrets or real-page fixtures belong in this repository.
|
Perkz Sell and manage access to private GitHub repositories. |
JustWrite Distraction-free writing sprints for macOS, with time or word-count goals. |
Releaseflow App updates and downloads. |
|
Passlock A macOS password manager with time locks and unlock challenges. |
||
|
Skillbox A self-hosted, versioned skills library for AI agents. |
Unclutter Remove page clutter with AI-powered, reusable browser rules. |
PageGrade Grade page clarity, writing, and on-page SEO. |
|
Council Let your coding agents deliberate together before making a plan. |
CodexMaxx Manage Codex accounts, usage, and active sessions on macOS. |
React Hanger A collection of useful React hooks. |
|
React Genie Animate React elements as they enter the viewport. |
MobX Router A simple router for MobX and React apps. |
|
All projects · GitHub · Follow on X · YouTube