I stopped ranking AI agent tools by total GitHub stars and started tracking star velocity instead. This week's #1 is a Codex "model routing" skill that's only 1 day old.

Most "top AI agent" lists are basically a snapshot of what got popular months ago. I wanted something that surfaces what's actually gaining momentum right now, so I rank projects by star velocity instead of total stars. The score looks at things like: 24h and 7d GitHub star growth percentile ranking within each source cohort time decay (so older spikes fade) cross-checking GitHub with the MCP registry The cross-check is important because it helps filter out one-off GitHub spikes and fake-star campaigns (the StarScout paper documented millions of fraudulent GitHub stars). A project that's exploding on one source but nowhere else gets down-weighted instead of jumping to the top. This week's fastest-rising agent tools: codex-model-routing-team — 107★, only 1 day old. It routes parallel Codex tasks to background workers with explicit per-worker models. Native subagents inherit the parent session model, so this gives you much finer control over cost and model selection. pilotfish — +359★. Multi-model orchestration for Claude Code. motion-anything — +354★. One thing I found interesting: the projects gaining traction aren't really about creating more agents. They're mostly focused on routing, coordination, and cost control across multiple agents. I published the full methodology (formula, weighting, decay, etc.), but I'll put the link in the comments so this post doesn't turn into a wall of text. Happy to answer questions or discuss the scoring if anyone's interested.