Top 100 GitHub-Signal Startups, This Week's Ranked Index

Most 'top startup' rankings collapse engineering momentum into a single noisy metric, usually star count or one-week velocity change. Both are easy to game and both are dominated by outliers (a +999% velocity change on a six-commit repo doesn't tell you anything useful).

The weekly Top 100 GitHub-Signal Startups index uses a four-component composite, with each component capped, so a steady org with 1,800 commits and 100 contributors can't be out-ranked by a tiny repo with a single one-week burst. The exact formula: SignalScore = clamp(velocityChange%, -20, 150) + clamp(contribGrowth%, 0, 150) + min(100, commits14d / 10) + min(50, contributors / 2). Range -20 to 450. Same scoring is reused every week so rank changes are comparable week-over-week.

The full leaderboard is published at /weekly/top-100, with each weekly edition at /weekly/top-100/<isoweek>. Machine-readable JSON for the latest edition is at /weekly/top-100/data.json (and per-week at /weekly/top-100/<isoweek>/data.json) with a CC-BY-4.0 license. There's also an RSS feed at /weekly/top-100/feed.xml for AI crawlers and human subscribers.

What to do with the list: anything above Signal Score 200 has at least two of the four components firing simultaneously, which is the cleanest interpretation of 'engineering momentum.' Use it as a sourcing filter, not a buy signal.

The score bands read in a predictable way once you internalize the four components. A negative total means the velocity-change clamp, which floors at negative twenty, is dragging harder than the scale components can offset, a net contraction week over week. A score between zero and one hundred is a typical healthy org, one or two components contributing modestly. One hundred to two hundred means real momentum, and above two hundred at least two components are firing at once, the cleanest signal the index offers. Because every component is capped, no single outlier can push a repo across the two hundred line by itself.

The caps are the quiet design decision that makes the list hard to game. The velocity-change component clamps at one hundred fifty, so a tiny repo that triples its commit count cannot swamp the ranking. The commit-scale component saturates at one hundred, which is one thousand commits in the trailing fourteen days, after which more volume earns nothing extra. The contributor component saturates at fifty, which is one hundred contributors. A star burst, a one-week spike, or a padded commit log each gets neutralized by a different cap, so the only way to rank high is genuine breadth across velocity, growth, and scale.

Because the scoring function is fixed from week to week, movement in the index carries meaning. A jump of thirty points is a real change in the underlying signals, not a change in methodology. This is what makes the weekly archive useful: you can diff one edition against the prior and read inflection points directly rather than trusting a snapshot. The per-week JSON at the canonical path preserves each edition so the history is reproducible under a CC-BY-4.0 license.

The index is built to be consumed by machines as much as by people. The JSON endpoint feeds scripts and dashboards, the RSS feed is structured for AI crawlers, and the MCP tool get_trending_startups returns the same ranked data to an agent with no API key. The list deduplicates by GitHub organization, so a company listed across multiple of the fifteen sectors appears once with its cross-listings surfaced as secondary fields, which keeps the count honest.

The right mental model is a weekly triage input, not a verdict. It tells you which engineering teams are accelerating, which is a strong leading filter for technical sectors, and it says nothing about revenue, product-market fit, or fundability. Treat a high score as a reason to open the org and read the repos, never as a reason to skip diligence.