aiagent.club

The question it answers

Most rankings show a single number — usually GitHub stars — at a single moment. Stars are attention, and attention is easy to manufacture and easy to hoard. A project can top a leaderboard for a year after people stopped using it.

The harder question is: what is actually being used, and what is gaining momentum right now? aiagent.club is built to answer that, by tracking many signals over time instead of one number at one instant.

Hype vs. real usage

For every project we hold attention signals (stars, watchers) next to real-usage signals (package downloads, model tokens processed, marketplace installs). When a project ranks far apart on the two, that gap is the story a static leaderboard hides — an "overhyped" star-heavy repo, or an "underrated" workhorse people install far more than they star.

Downloads and tokens are not un-gameable, but they are much harder and more expensive to fake than a star, and they track how the tool is actually pulled into real projects.

Momentum over size

Cumulative totals reward whoever started first. We compute trailing-window deltas — 7-day star gains, week-over-week download surges, new releases — and surface them in Movers and the Changelog. That is where you see a project break out weeks before it reaches the top of any all-time list.

Sources & metrics

Signals are collected from public sources, both global and Chinese:

• GitHub — stars, forks, commits, releases, open issues/PRs (frameworks, MCP servers, agent tooling)
• npm & PyPI — weekly / monthly package downloads (real usage of the libraries)
• OpenRouter — tokens processed per day per model (the hardest-to-fake model-usage signal)
• ModelScope — download counts for Chinese models (Qwen / DeepSeek / GLM …)
• VS Code Marketplace — install counts for coding agents

Project descriptions and metadata come from each source of record (GitHub, npm, PyPI). Data is shown in its source language and never machine-translated — translating a name or license is a bug you can't detect later.

Resisting the games

A ranking is only trustworthy if it fights the ways people inflate it:

• Star-farm detection — repositories accumulating stars implausibly fast (sustained hundreds of new stars per day on already-large repos) are flagged and excluded, so bought-star projects don't crowd out real ones.
• Cross-validation — no single gameable metric decides a project's standing; attention and usage are shown side by side, so a number that has been inflated stands out instead of winning.
• Canonical de-duplication — when a repository is renamed or moved between orgs, its old and new identities are merged into one canonical entry so a project isn't double-counted.

Cadence & data model

Every tracked metric is snapshotted twice a day. History is append-only — we record each day's value and never rewrite the past; that accumulating time series is the whole point. Collection is idempotent and self-healing: a rate-limited or failed source is retried on the next run and never corrupts what was already stored. Where a public history existed (e.g. a repo's full star history), we backfilled it once so trends start deep rather than from zero.

Public vs. private, and caveats

Rankings, current snapshots, and trend charts are public; the full day-by-day history is kept private. The collectors are open source. A few honest limitations: newly discovered projects start as a single data point and grow a trend one day at a time; download figures are rolling windows, not lifetime totals; and reconstructed star history reflects stars that still exist today, not a past instantaneous peak.