Catalogue data

The ten most-starred AI tools and the ten most-installed share one name

Across the 78 tools where we measure both, star rank and install rank are nearly independent — Spearman 0.131. TensorFlow is first by stars and thirteenth by installs. LangChain is the only name in both top tens.

Updated 21 October 2026·7 min read·No paywall

A dazzling shop front covered in stars with an admiring crowd outside and an empty doorway, next to a plain green shop with a queue of people walking out carrying parcels — Admired on the left, installed on the right — and the two rankings share one name in ten.

TL;DR — the short version

The ten most-starred tools and the ten most-installed share exactly one name, LangChain. Across the 78 tools where we measure both, star rank and install rank are nearly independent — Spearman 0.131, so the star order explains 1.7 % of the install order.

There is no exchange rate. In the least-installed quarter of the cohort a star is worth 0.56 downloads a month; in the most-installed quarter, 607. The same star, 1,084× apart.

Star velocity is no better. This week’s star gain — the signal every trending page is built on — correlates with installs at 0.082.

What follows for you: a star count answers “did this impress someone once?”. It does not answer “is anyone running this?”. For a dependency, find the second number.

Every tool directory on the internet, ours included, sorts by stars. It is the only number every repository carries, so it became the number everyone uses. We can now check what it is worth, because for part of our catalogue we also measure something harder to fake: how many times the package is actually installed.

78 of the 242 tools we track publish a PyPI package, so we have both figures for them. The two rankings barely agree.

The most-starred tool is thirteenth by installs

Tool	Stars	Rank by installs	Installs per day
TensorFlow	201,193	13^th of 78	391,010
Firecrawl	193,164	25^th	175,424
Open WebUI	156,214	41^st	42,803
Langflow	155,855	64^th	1,240
LangChain	148,096	2^nd	5,599,795
Browser Use	118,573	20^th	264,848
Whisper	110,897	27^th	163,803
vLLM	94,167	38^th	61,019
RAGFlow	92,501	67^th	536
PaddleOCR	91,397	45^th	40,890

Ten tools at the very top of the star ranking, and their install ranks run from 2^nd to 67^th. Only LangChain sits near the top of both. Move the whole cohort from one ranking to the other and the median tool shifts 23 places out of 78; a third of them — 26 tools — move more than 30.

There is no exchange rate between a star and an install

If stars were a weak but honest proxy, you would expect a roughly constant ratio: so many downloads per star, give or take. Sort the cohort into quarters by install volume and the ratio does not drift — it changes by three orders of magnitude.

The median tool in the least-installed quarter sees 0.56 monthly downloads for each star it carries; in the most-installed quarter, 607. A star in one part of the catalogue is worth 1,084× a star in another, which is another way of saying the unit means nothing on its own.

The widest gap we found

MetaGPT carries 70,932 stars, 3,087 more than scikit-learn. scikit-learn is installed 184,326,320 times a month; MetaGPT, 3,393. That is a factor of 54,325 in the opposite direction to the applause.

At the far end sits BabyAGI: 22,356 stars, 101 installs a month. Three a day. It was the repository everyone linked to in the first weeks of the agent wave, and the star count has kept it visible ever since, because stars never expire.

Where this measure is unfair, and we should say so

Nine tools in the cohort hold 20,000+ stars and see fewer than 1,000 installs a day. It would be easy — and wrong — to present all nine as repositories nobody uses. For several of them, pip is simply not how you get the software: RAGFlow and Label Studio are deployed with Docker, InvokeAI ships its own installer, LLaMA-Factory is cloned rather than installed, and fastbook is the notebook set of a book. Their PyPI figure understates them, and we have no equivalent number for the channel they actually use.

The finding survives that correction, because it does not rest on those nine. It rests on 78 pairs of numbers whose rank correlation is 0.131, on a top ten that shares one name out of ten, and on the fact that the same unit buys 0.56 downloads here and 607 there.

The deeper limit is ours. Only 87 of our 242 tools — 36 % — have any install figure at all, because the rest ship as binaries, desktop apps, editor extensions or container images we do not count. Ollama, llama.cpp and ComfyUI are among them. For nearly two thirds of the catalogue, the star count is still the only number on the card, and this article is the argument for fixing that rather than for trusting it.

And trending is worse than stars

If a total star count is a poor proxy, perhaps recent momentum is better — the signal behind every “trending” page. It is not. Across the 72 tools where we have both, the star gain of the past seven days correlates with monthly installs at 0.082: lower than the total, and indistinguishable from noise.

That makes sense once you say it plainly. A star is given by someone reading; an install is performed by someone building. The first is a reaction to a README, the second a decision in a requirements file. We have been ranking the second by measuring the first.

Before you depend on something, find the second number. For a Python package, pypistats or the PyPI page gives monthly downloads in one click. For a container, Docker Hub shows pulls. For anything else, the release download counts on the repository. If the project has no second number at all, treat the star count as what it is — evidence that it was once interesting — and look for whether anyone is still pushing code, which our health tracker measures.

Method

Figures come from the olud.ai tool metrics, snapshot of 20 October 2026, covering 242 tools. The cohort is the 78 tools that publish a PyPI package and carry a star count; Docker Hub (15 tools) and npm (14 tools) are measured separately and give 0.143 and 0.037, which we report as corroboration, not as independent proof — those cohorts are far too small to carry a correlation. Correlations are Spearman rank coefficients, which compare orderings rather than values, so a single giant cannot distort them. “Installs per day” is the monthly figure divided by 30.44. A tool without a PyPI package has no download figure and is excluded, never counted as zero — that is why the cohort is 78 and not 242. Download counts are not people: PyPI totals include continuous-integration runs, mirrors, container builds and transitive dependency resolution, which inflates mature libraries pulled in by other packages — part of the gap we measure is machine traffic. One snapshot, no time series: we cannot say whether this gap is widening.

Data: olud.ai tool metrics, refreshed daily — method on /methodology.html.

Keep reading

One fork for every 6.9 stars — the other counter on the page, and what it reveals about what a repository is.
459 repos hold half the stars — how steep the applause curve is in the first place.
127 repos above 10,000 stars are held up by fewer than ten people — the third number worth reading before you depend on anything.