Krakt Radar: AI launches from GitHub, Hugging Face, Hacker News & arXiv

One run gives you a single, deduplicated feed of what's new and moving in AI:

• GitHub: repositories from github.com/trending, plus new AI repos from the GitHub Search API, filtered by AI keywords
• Hugging Face: trending models and Spaces (official Hub API)
• Hacker News: "Show HN" posts that match AI keywords (official Algolia HN API)
• arXiv: newest papers in cs.AI, cs.LG and cs.CL (official arXiv API, with the requested 3-second delay between calls)

Every item gets a momentum score (0-100). Items that show up on more than one source are linked, so you can see when a project launches on GitHub, gets posted to HN and trends on Hugging Face at the same time.

Who it's for

AI newsletter writers, VCs and scouts, devrel and competitive-intel teams, and anyone building an "AI news" agent or dashboard. Schedule it daily with Only new items on and you get just the new items each day.

Input

Field	Default	Notes
sources	all four	github, huggingface, hackernews, arxiv
period	daily	daily, weekly, monthly
keywords	built-in AI list	Applied to GitHub and HN. If you set custom keywords, they apply to all sources.
maxItems	100	Total items returned (each one is billed)
maxItemsPerSource	50	Fetch cap per source
onlyNew	false	Skips items returned by earlier runs that used the same stateKey
stateKey	default	Use one key per schedule
minMomentumScore	0	Drops items scoring below this
breakoutAlerts	true	Marks multi-source items as breakouts (billed as an extra event)
githubToken	–	Optional. Raises the GitHub Search API rate limit.

Output (one row per item)

source, kind (repository/model/space/show_hn/paper), id, title, url, description, author, metricName, metric (stars/likes/points/authors), metricDelta (GitHub stars in period, HF trending score), metricChangeSinceLastRun, publishedAt, tags, momentumScore, crossSourceHits, crossSources, isBreakout, isNew, scrapedAt. Source-specific extras: downloads (HF), comments and hnUrl (HN), codeUrl (arXiv).

A RUN_SUMMARY record in the default key-value store shows, for each source, whether it succeeded, the method used and the item counts.

How the momentum score works

Within each source, items are ranked by that source's own signal:

• GitHub: stars gained in the period
• Hugging Face: trending score
• Hacker News: points + 0.5 × comments
• arXiv: recency

The rank is scaled to 0-85 (0-60 for arXiv, which has no engagement signal). Each additional source the item appears on adds +15. An item is a breakout when it appears on 2+ sources and scores ≥ 65. Cross-source matching uses repo and model names and the links in posts and abstracts. It is heuristic and doesn't happen on most days.

Pricing

Pay per event:

• Item: $0.002 per item returned
• Breakout: $0.01 extra per item flagged as a breakout (turn off with breakoutAlerts: false)

The Actor respects your maximum cost per run and stops cleanly when it reaches it.

Reliability and limits

• If one source fails, the others still return data. The failure is recorded in RUN_SUMMARY and in the status message.
• GitHub trending is read from the public HTML page. If its layout changes, the Actor falls back to the GitHub Search API automatically.
• Uses only public data and official APIs where they exist. No logins, no proxies.
• Hugging Face Spaces have no description field, so their description is the SDK name.
• arXiv publishes in daily batches. On weekends, the "daily" window falls back to the latest batch.

Data belongs to its respective sources. Follow GitHub's, Hugging Face's, Y Combinator's (HN) and arXiv's terms when you reuse it.