AI Radar

A live dashboard of the latest news from the leading AI companies, aggregated from their official and community feeds into one page. AI Radar is a static site published with GitHub Pages and refreshed by a scheduled GitHub Action.

Live: https://ai-radar.eu/

What it does

AI Radar follows 13 companies (OpenAI, Anthropic, Google AI & DeepMind, Meta AI, Microsoft AI, NVIDIA, Hugging Face, xAI, Perplexity, Mistral, Cohere, DeepSeek, Moonshot AI (Kimi)), plus an Across AI card for industry news that is not tied to a single company (policy, funding, society), fed by a Google News query and MIT Technology Review with optional hand-picked items from curated.json. In total the build reads 42 RSS and Atom feeds: 40 for the companies and 2 for Across AI. The dashboard brings the updates together as:

• A Today in AI carousel with up to five top stories. By default they are picked automatically and labelled Auto-selected: recent model releases first, one story per company, only posts with an image, from the last 72 hours (widened to 7 days on a quiet day). The carousel advances every 7 seconds and pauses on hover, on keyboard focus, in a background tab or with motion turned off; the arrows, the progress segments, the arrow keys and a swipe move between slides. With the optional digest (see below) it shows the AI-written briefing instead.
• A Latest column with the five newest posts.
• A live status line in the header: when the data was last updated, a countdown to the next scheduled build and the number of sources. The page checks data.json every five minutes and when the countdown runs out. New posts wait behind a "↑ N new posts" pill, so nothing moves while you read; if a check fails, the status switches to Offline and a bar offers Retry.
• By company: one card per company, ordered by the number of posts this week, then by the newest post. A card shows its latest post plus three more (two on mobile) and expands to at most 12, with an "All N in Timeline →" link when there are more. Across AI posts show the logo of the company they are about.
• Timeline: every post in chronological order, grouped by day with a post count per day, 50 at a time with "Load more".
• Category chips (Model releases, Products, Research, Developer, Hardware, Applied AI, Business, Safety & policy, Other), free-text search ("/" jumps to the field) and saved posts: bookmark any post and show only those with "Saved · N". The view and category are kept in the URL (for example ?view=timeline&cat=model-releases), so a filtered view can be shared.
• Motion and theme toggles. Both follow the system setting (reduced motion, light or dark) until you choose, and the choice carries over to the about, contact, privacy and disclaimer pages.

All times and the day grouping are in UTC, like the build schedule. No analytics, no tracking, no cookies, no web fonts and zero third-party requests: text uses the system fonts, logos are inline and images are served from the site itself. The only data stored is kept in the browser's localStorage under airadar-theme, airadar-motion and airadar-saved (the ids of saved posts). Keys from the previous version (mm-theme, mm-saved-v1 and the other mm-* keys, ai-dashboard-cache-v3) are migrated or removed on the first visit.

Model History

ai-radar.eu/models/ is a source-based history of the models of the same 13 companies: which model an organisation released, when, what changed, and how the models relate (successor, revision, variant, derived from). Every fact links to its sources, dates keep the precision the source gives, and each source shows whether the original is still online, archived or gone. It is not a benchmark or a ranking.

• An Explorer with search, filters (organisation, family, category, year, status, evidence, open weights) and sorting; a page per organisation with its chronology and families; a page and a timeline per model.
• The pages are pre-rendered static HTML generated by scripts/model-history/build.mjs from the curated JSON in model-history/ (see model-history/README.md for the data rules). The Model History workflow validates every pull request and regenerates models/ and sitemap-models.xml on main; a daily source-health workflow checks every source address. AI Radar posts are linked to models automatically by their aliases.
• New models are added automatically, without anything running on a local machine: the Model History build opens a new-model issue for every model name in AI Radar titles that no record knows; a daily Claude routine in Anthropic's cloud researches those names and pushes the new records to a claude/ branch; and model-history-auto.yml publishes them after the same checks as a pull request (validation, tests, validation after news matching). These entries are not reviewed by a person before they go live; the About page says so. Details in model-history/README.md.
• The same privacy rules as the dashboard: same-origin static files, no new storage keys, the search term never leaves the page. The data is licensed CC BY 4.0 (model-history/LICENSE.md).

How it works

A GitHub Action (.github/workflows/build-feed.yml running scripts/build-feed.mjs) runs every two hours (minute 17 of every even UTC hour) and writes three files to the repository root:

• data.json — structured items (title, link, date, summary, image, company, source) plus the build schedule; the dashboard's only data source.
• feed.xml — a combined RSS feed used for the email subscription via Blogtrottr.
• digest.json — the optional daily briefing (only when enabled; see below).

The dashboard (index.html with assets/app.js) reads data.json from its own domain. There is no fallback to live feeds or proxies: if the file cannot be loaded, the page says so and offers Retry and the RSS feed. Missing thumbnails and summaries are backfilled from each article's Open Graph tags during the build, cached across runs. Article images are downloaded, resized and committed to img/, so visitors' browsers make no third-party requests.

Besides items, data.json carries generated and lastFetchedAt (build time), nextFetchAt (the next scheduled build, which drives the countdown), intervalMinutes (120), sources (the number of source feeds, shown as "N sources") and digest (true only when a usable digest.json exists). Each item has a stable id, a hash of its normalised link that bookmarks and digest items refer to, and Across AI items that name a tracked company get about. The dashboard uses about as the build wrote it: an item with an id but no about is not about a tracked company. The new fields are additive: the dashboard also works with an older data.json without them. It then computes id and about itself, with a copy of the build's rules (ABOUT_RULES in scripts/build-feed.mjs and assets/app.js must stay identical), shows the card count as the number of sources and derives the next refresh from the cron schedule.

Daily briefing (optional)

The briefing is optional and the site does not need it. Without an API key no digest.json is written, data.json says "digest": false, the page never requests the file, and the carousel auto-selects its stories as described above.

When the repository secret ANTHROPIC_API_KEY is set, the build generates a "Today in AI" briefing once per UTC day with the Anthropic Messages API (Claude Haiku), writes it to digest.json, sets "digest": true in data.json, and prepends the briefing to feed.xml for email subscribers. The digest uses format v2: 3 to 5 items, each with a title, one or two sentences of text and the sourcePostId of the data.json post it is based on, so the carousel can show that post's image, company and link. Items that point at an unknown post are dropped, and a result with fewer than three valid items keeps the previous digest; a legacy v1 digest is replaced on the next run. The page shows a digest from today or yesterday (UTC), labelled AI-generated, and otherwise falls back to the auto-selected stories. Without the secret the feature stays dormant and the rest of the build is unaffected.

To enable it: add the secret under Settings → Secrets and variables → Actions, then run the workflow (it also runs on the two-hour schedule).

Local preview

The site is plain HTML, CSS and JavaScript with no build step and no dependencies: index.html with assets/tokens.css, assets/app.css and assets/app.js for the dashboard, and assets/pages.css and assets/pages.js for the about, contact, privacy and disclaimer pages. Serve the repository root with any static web server and open /. The pages use root-relative paths (/data.json, /assets/…), so opening index.html straight from disk does not work, and the internal links are extensionless (/about), which GitHub Pages resolves to about.html. In the browser console, window.__airadar.poll() checks for a new build right away. The build script only runs in CI and needs Node 20+. ImageMagick is optional: when it is installed (as in CI), the build resizes the downloaded article images; without it, the originals are kept.

Versioning and releases

The project follows Semantic Versioning, and every notable change is recorded in CHANGELOG.md.

Tagged releases (vX.Y.Z) automatically create a GitHub Release: pushing a tag runs .github/workflows/release.yml, which generates release notes and a "Full Changelog" link comparing the tag to the previous release. So each release shows exactly what changed since the last version and links back through the whole history. Routine feed-data commits made by the build bot are excluded from the notes.

To make per-change history easy to review, prefer pull requests over direct pushes: create a branch, open a PR (a template is provided), review the diff, then merge. Merged PRs are grouped into the next release's notes.

# cut a release once changes are on main
git tag -a v2.0.0 -m "AI Radar 2.0.0"
git push origin v2.0.0

Sources and disclaimer

AI Radar reads publicly available feeds and displays them unchanged. Some companies are read from an official feed; the others from a community feed that may lag behind. No rights can be derived from the information shown, and no guarantee is made as to its accuracy, completeness or timeliness. All rights to the posts remain with the original sources. Brand names and logos are the property of their respective owners and are used for identification only; this project is not affiliated with or endorsed by any of the companies listed.

License

The code in this repository is released under the MIT License (see LICENSE). The license covers the code only, not the aggregated third-party content.

Made by Benjamin Nieuwenhuijzen.