How momentum is scored

Cresting ranks tools by how fast they are rising, not how big they already are. The formula below is the whole story — the same math for every tool, recomputed hourly, with no manual edits and no way to buy a better number.

The formula

velocity   = weighted mean of normalized signals
             (0.5·stars_7d + 0.2·stars_24h + …, weights renormalize
              over the signals a tool actually has)

recency    = exp(−age_days / 14)          # brand-new gets a lift
decay      = exp(−days_since_signal / 30) # goes cold if traction stalls
corrob     = 1 + 0.15·(sources − 1)       # capped at +45%
quality    = 0 or 1                       # spam/dead/NSFW never rank

momentum   = 100 · velocity · (0.6 + 0.4·recency) · decay · corrob · quality

Each raw signal is log-compressed, then percentile-ranked within its own source cohort over a trailing 30-day window — so a 500-star week from a small project and a 100k-star week from a giant both max out at the top of their cohorts. Velocity beats size, by construction.

Signals we read today

• GitHub star velocity (24h and 7d windows)
• Official MCP Registry presence and freshness
• More sources (Product Hunt, Hugging Face) are being wired in; each one strengthens corroboration.

Resisting hype

Sudden bursts wildly out of line with a tool's own history get their contribution capped (the score shows a "clamped" marker in its breakdown). Percentile normalization means no single whale can distort the scale, and decay means yesterday's spike doesn't coast.