Vibecode Weekly

A GitHub Pages site tracking vibe coding trends: weekly growth signals (GitHub, Hacker News, Google Trends, VS Code Marketplace installs) plus a daily feed of news and YouTube videos, from sources configured from a simple Google Sheet that admins can edit directly, and relevant products tracked from Product Hunt. No backend — everything renders client-side against static CSV/JSON files committed by two scheduled GitHub Actions workflows.

How it fits together

Google Sheet (non-technical source/topic config - feeds the daily
workflow only; the weekly workflow's tool lists are hardcoded in
scripts/fetch_trends.py, not Sheet-driven)
        |
        v
GitHub Actions (daily, 05:00 UTC)              GitHub Actions (weekly, Mon 06:00 UTC)
  -> scripts/fetch_news.py                       -> scripts/fetch_trends.py
  -> scripts/fetch_youtube.py                     -> scripts/compute_signals.py
  -> scripts/generate_digest.py (optional)        -> scripts/fetch_producthunt.py (optional)
  -> writes data/news.json, data/youtube.json     -> writes data/*.csv, data/signals.json,
  -> writes/updates data/status.json,                data/status.json, data/producthunt*.json
     data/digest.json (optional)
  -> commits + pushes                             -> commits + pushes
                    |                                       |
                    +-------------------+-------------------+
                                        v
                          index.html / leaderboard.html / compare.html
                        (fetches the JSON/CSV directly, no build step)
                                        |
                                        v
                                   GitHub Pages

Repo layout

• index.html, leaderboard.html, compare.html — the site
• assets/ — shared CSS/JS (style.css, app.js, trend-chart.js, radar-chart.js, tool-profiles.js)
• data/ — the CSVs/JSON both workflows write to, and what the site reads
• scripts/ — the six Python scripts (trends, news, YouTube, signals, digest, Product Hunt)
• .github/workflows/ — the two scheduled workflows
• docs/GOOGLE_SHEET_SCHEMA.md — the exact columns the Sheet needs

Non-technical source management

News and video sources live in a Google Sheet, not in code. See docs/GOOGLE_SHEET_SCHEMA.md for setup and the exact column format. Adding a source is just a spreadsheet row; no code change or redeploy needed.

Required secrets

Set these under Settings → Secrets and variables → Actions:

Secret	Used by	Notes
GITHUB_TOKEN	weekly workflow	auto-provided, no setup needed
TRENDS_MCP_API_KEY	weekly workflow	see the Google Trends caveat below
SHEET_CSV_URL	daily workflow	the published-CSV URL of the sources Sheet
YOUTUBE_API_KEY	daily workflow	YouTube Data API v3 key, free tier
GEMINI_API_KEY	daily workflow	optional. Gemini free-tier key - adds an LLM second opinion on top of the keyword-based relevance/category checks in fetch_news.py and fetch_youtube.py (both run keyword-only if unset, capped at 30 calls/run each - see MAX_GEMINI_CALLS_PER_RUN), and powers the auto-written digest in generate_digest.py. That script runs daily but self-limits to writing a new digest roughly once a week (MIN_DAYS_BETWEEN_DIGESTS); skipped entirely, leaving any prior digest in place, if unset
PRODUCTHUNT_API_KEY	weekly workflow	A non-expiring developer_token from your Product Hunt account's API dashboard (producthunt.com/v2/oauth/applications) - see the Product Hunt section below. Skipped (not a failure) if unset

The VS Code Marketplace signal (fetch_vscode_marketplace_installs() in scripts/fetch_trends.py) needs no secret at all - it's an unauthenticated public endpoint, so there's nothing to add here for it.

What each data source measures

GitHub — repo count per configured topic tag. Official API, no auth required (a token raises the rate limit).

Hacker News — story and comment mentions per search term, last 7 days, via the official Algolia HN Search API. No auth needed.

Google Trends — relative search interest (0–100), via Trends MCP (trendsmcp.ai), an unofficial third-party proxy, not a Google product. Free tier caps at 100 requests/month, 20/day — this is why the term list stays short. Treat this as the least stable link in the pipeline; a manual CSV export or a paid alternative (SerpApi, Apify) are the fallbacks if it breaks. See scripts/fetch_trends.py for the parsing details, since the response schema is inferred, not officially documented.

VS Code Marketplace — total install count per extension, via the Marketplace's own public Extension Gallery API (marketplace.visualstudio.com/_apis/public/gallery/extensionquery). No auth or API key required, and it's undocumented for third-party use but is Microsoft's own endpoint (the same one the VS Code client itself queries), queried by exact extension ID (VSCODE_EXTENSIONS in scripts/fetch_trends.py). This is the one signal here that's a direct usage count rather than a proxy for one — but it only exists for tools that ship an actual VS Code extension (Cursor and Windsurf are standalone forked editors, not extensions; Replit Agent, Devin, and Lovable are browser-only, so they have no extension to count installs for either).

News (RSS) — headline, link, date per configured feed. Feeds only, not scraping — scraping is fragile and often against a site's terms of service.

YouTube — long-form videos (last 60 days) per configured search topic, ranked by views-per-day-since-published as a proxy for "gaining popularity," since there's no official trending-by-niche endpoint. Search results are ordered by view count (not date) within that window so genuinely popular videos aren't crowded out by newer, less-watched ones before stats are even fetched. Shorts (<=3 minutes, YouTube's current eligibility threshold) are excluded - see SHORTS_MAX_SECONDS in fetch_youtube.py.

Product Hunt — recent launches from Product Hunt's own "vibe-coding" topic, via their v2 GraphQL API (api.producthunt.com/v2/api/graphql), authenticated with a developer_token (see PRODUCTHUNT_API_KEY above) - no OAuth flow needed for a script like this one. Unlike every other source here, this one isn't tied to a pre-configured list of tools at all: it's this project's answer to "how would we notice new and growing tools before someone manually adds it to TOOLS/GITHUB_TOPICS/etc." - see scripts/fetch_producthunt.py's docstring. The site cross-references each launch's name against assets/tool-profiles.js client-side and flags anything not already tracked. The same script also writes data/producthunt_top.json and data/producthunt_reviewed.json - always-fresh (not accumulating) snapshots of the all-time top-voted and top-reviewed posts within that same "vibe-coding" topic, for the site's "Top voted"/"Most reviewed" lists. Both are derived from one shared, broad fetch (not two separate order-specific queries), so a post with few votes but a genuinely high review count still gets found - see the module docstring for why that distinction matters. producthunt_reviewed.json only includes posts with at least one real written review (a separate, less common action than a vote), not every post sorted with 0-review ones at the bottom.

All three files are scoped to the vibe-coding topic alone - Product Hunt's own "Best vibe coding tools" page draws from a separate, editorial Category grouping that isn't reachable through the public API, so established tools that don't carry the vibe-coding topic tag themselves (Cursor, Lovable, Windsurf, and similar) won't appear in these three lists. Two terms from Product Hunt's own API docs worth knowing if this data is ever used beyond this proof of concept: it must not be used for commercial purposes without contacting them, and they ask for attribution linking back to Product Hunt (see each page's footer disclosure).

Signals, not a score

scripts/compute_signals.py writes each tool's four signals as separate fields in data/signals.json - honestly null where a signal genuinely isn't tracked for that tool - and does not rank tools against each other or blend them into a composite score. Combining narrow, noisy proxies (GitHub topic-tag counts, comment volume on one forum, search interest via an unofficial third-party proxy) into a single number doesn't make the result more accurate, and it makes it harder to audit: a reader can't tell whether a blended score change reflects real momentum or one of those proxies' known failure modes, and any weighting between them would be arbitrary rather than validated. leaderboard.html shows the four signals as a sortable table (click a column to sort by it); compare.html's radar chart plots them as four independent axes.

Not every tool gets all four signals. Cursor has no Trends coverage ("Cursor" is too generic a search term to track cleanly). Replit Agent, Devin, and Lovable have no GitHub signal at all - none of them are tools people build a public GitHub ecosystem of extensions/example repos around the way an IDE or CLI tool is, so a repo count would be thin (Replit Agent/Lovable - usage mostly stays on the vendor's own hosted platform) or actively misleading (Devin - "devin" is also just a common first name). Only Claude Code, Codex, and GitHub Copilot have a VS Code install count - Cursor and Windsurf are standalone forked editors rather than VS Code extensions, and Replit Agent, Devin, and Lovable are browser-only. A missing signal shows as a dash on the site, never a fabricated zero.

Adding a tool

Edit the TOOLS dict at the top of scripts/compute_signals.py — map a display name to the matching metric names already being tracked in the two CSVs. Set "github", "trends", or "vscode" to None if that signal genuinely isn't trackable for this tool (see "Signals, not a score" above)

• "hn" is the one every tool is expected to have, since Hacker News needs no topic tag, Trends term, or Marketplace listing to configure first. If the tool isn't tracked yet, add it to the GITHUB_TOPICS/HN_TERMS/TRENDS_TERMS lists in scripts/fetch_trends.py first (skip GITHUB_TOPICS if you're setting "github": None), and to VSCODE_EXTENSIONS (same file) with its exact Marketplace extension ID if it ships a VS Code extension. Also add it to assets/tool-profiles.js (compare.html's hand-curated facts) and to TOOL_MATCHERS in compare.html if you want it picked up by the "recent mentions" count there.

Hosting

The repo can stay private. Recommended: connect Cloudflare Pages (or Netlify/Vercel) directly to this repo — works on the free tier without requiring the repo to be public, and without depending on the org's GitHub plan. If deploying via GitHub Pages instead, note that private-repo Pages requires GitHub Team or higher; on Team+ you can keep the repo private while setting the published site to public visibility.