GitHub Star Velocity: 30 Open-Source AI Agents Measured 7 Days Apart

<a href="https://theagentsindex.com/blog/github-star-velocity-open-source-ai-agents-2026/" target="_blank" rel="nofollow noopener"><img src="https://theagentsindex.com/charts/github-star-velocity-open-source-ai-agents-2026.png" alt="Commit activity among 32 tracked open-source repos — The Agents Index" title="Commit activity among 32 tracked open-source repos — The Agents Index" width="640" loading="lazy" style="max-width:100%;height:auto" /></a>

Of the 94 AI agent products published in our index, all 94 have a determined open-source status and 33 of those are open source. Of those 33, 32 have a public GitHub repository we could identify and query; the 33rd, Vellum, states it is open source but names no repository we could confirm, so it is out of scope for a page that measures repository activity specifically. On 2026-08-18 we queried the GitHub REST API for the star count, fork count and last-push timestamp of every one we could then identify. On 2026-08-25 we did it again, the same way, from the same API. This page is the second measurement set against the first, seven days apart for the 27 repositories with an 2026-08-18 baseline, and against each repository’s own dated first count for the rest.

Two dated counts is the whole point. A single star count is a number anyone can fetch in a second, and it tells you almost nothing on its own. The rate of change needs two measurements taken by the same method at two known dates, and for 27 of these tools the gap is exactly 7 days: 2026-08-18 to 2026-08-25.

Two of last edition’s three quiet repos have since shipped

The most decision-relevant finding last edition was three tools whose default branch had received no push at all, gaining stars anyway. Measured again this week, that finding has partly reversed, which is exactly the kind of thing a snapshot claim should be checked against rather than repeated:

GPT Researcher, the third repository flagged last edition at 31 days quiet, pushed a commit on 2026-08-24 and drops off this list. Its own maintainers ended the gap before we measured again, which is the reason this page carries a date rather than a permanent verdict.

Cartesia’s Line SDK, the counter-example last edition at 29 days quiet and zero stars gained, also shipped since: it pushed on 2026-08-23 and picked up 2 stars over its own 20-day measurement window, still nearly flat but no longer the clean zero it was.

Between the two still-quiet repositories, AutoGen and Aider, that is 312 stars for zero commits over the periods measured. AutoGen at 132 days is consistent with the Microsoft-declared maintenance mode already documented on its listing. Aider at 95 days remains a genuine, growing gap with no announced explanation we found.

One caveat we would rather state than have someone find. pushed_at reports the default branch only. A project can be actively maintained in forks, on release branches or by a maintainer who batches merges, and none of that shows up here.

The leaderboard: OpenAI Codex CLI led on both counts this edition

Stars gained per day for the 27 tools with a prior dated count. Every baseline in this table is our own published figure from an earlier count, not a number lifted from anywhere else.

OpenAI Codex CLI gained 11,317 stars in 7 days, more than four times the next tool on the list. At 1,616.7 stars a day it is adding roughly OpenClaw’s entire prior-week gain every four hours. That is a large enough jump to flag rather than just report: it lands the same week GPT-5.4 and GPT-5.4 mini were scheduled to retire from ChatGPT-signed Codex sessions (2026-08-31), and a rust-v0.149.0 release on 2026-08-20 shipped a new agents dashboard, which together are the kind of events that move a repository this fast. We have not independently confirmed a single cause and are stating the coincidence rather than a claim.

OpenCode, last edition’s leader at 406.5 stars a day, slowed to 365.6 but still holds second place comfortably ahead of n8n.

OpenClaw, Goose, LangGraph and AgentsKit graduate into this table for the first time this edition: each had only a single 2026-08-18 data point last time, with no prior figure to compare against. All four now have a real rate.

Fastest growth is a different question from most growth, except at the very top

Rank the same repositories by percentage gained instead of absolute stars, and this edition the two questions briefly agree at the top before diverging again. OpenAI Codex CLI grew 10.61% in 7 days, the fastest relative growth of any repository in this set, ahead of Microsoft Agent Framework’s 1.83% and OpenCode’s 1.29%.

Below the top, the pattern from prior editions holds: a project adding a few hundred stars off a base in the low thousands, like Microsoft Agent Framework’s 12,869, is being discovered right now, while a project adding thousands off a base in the hundreds of thousands, like n8n’s 201,077, has already been discovered and its absolute number mostly reflects how long it has been around.

Two tools have their first tracked figure, so they get no rate yet

Kilo Code and Sim both joined the index’s open-source set recently and are getting their first GitHub snapshot from this census. Inventing a rate from a single measurement is the one thing this method cannot do.

Both will get a real stars/day figure next edition, measured against these numbers as the new baseline.

A note on two organization moves

Two repositories tracked here now resolve to a different GitHub organization than last edition, and both moves are worth naming rather than passing over silently. Aider’s repository has moved from paul-gauthier/aider to Aider-AI/aider; GitHub’s redirect resolved the request automatically and the star count is continuous. Goose’s repository now resolves to aaif-goose/goose rather than block/goose, consistent with its stewardship by the Linux Foundation’s Agentic AI Foundation rather than Block directly; the star count is likewise continuous through the redirect.

Every listing in the index now has a determined open-source status

29 of the 32 tracked repositories pushed a commit within three days of this count. The quiet ones, Aider, Microsoft AutoGen and AgentsKit (5 days), are the exception here, not the pattern, and only the first two are quiet enough to be a real finding rather than an ordinary gap between releases.

All 94 published product listings now carry a researched open-source status: 33 are open source and 61 are not, with no unknown bucket left. MindStudio, the one listing whose license was undetermined as of the last census, has since been researched and come back closed source. Since then, Clay left the index (it was closed source, so its departure doesn’t move this count), and Vellum joined as open source without a confirmed repository, which is why the population above reads 33 open source but only 32 tracked here. Since this page was last checked, Lead Scorer also joined the index as a closed-source SaaS product, moving the published total from 93 to 94 without changing either the open-source count or the trackable-repository count.

For license terms rather than activity, our open-source license census is the reference. It is worth reading alongside this page, because several of the largest repositories here carry a licensing caveat a star count will not show you: n8n’s fair-code license and Dify’s modified Apache terms are both source-available rather than OSI-approved, and Sim’s and Vellum’s open-source claims are sourced without a confirmed license name.

What this measures, and what it doesn’t

A star is a free, one-click action. It measures interest, not code quality, security posture, test coverage or whether your team should build on the thing. We publish velocity because it is a genuinely different signal from a raw count, not because it ranks anything.

The one thing it does well is catch the divergence in the first table, and catch when that divergence stops being true. When a project’s stars climb while its commits stop, the star count is describing its past and the commit log is describing its present, and this edition shows that gap can close between two measurements. Use this page to see which of those two you are looking at, then use each tool’s own listing for pricing, licensing and the sourced verdict before you decide anything.

Method

Every figure marked 2026-08-25 was read from the GitHub REST API (api.github.com/repos/{owner}/{repo}) between roughly 16:20 and 17:20 UTC on 2026-08-25: stargazers_count, forks_count and pushed_at, which we report as “last push”. Every prior-count figure is our own previously published number, read the same way on its own date: 2026-08-18 for 27 repositories, and an individual dated baseline (2026-08-02, 2026-08-05 or 2026-08-06) for the three that joined the index between the two main counting passes.

Repository identity came from each listing’s own sourced research where we hold it, and from the same two named exceptions carried forward from the prior edition: OpenCode is anomalyco/opencode, not the archived opencode-ai/opencode that shares its name, and OpenHands resolves through a permanent redirect from All-Hands-AI/OpenHands to OpenHands/OpenHands after an organization rename. This edition adds two more: Aider now resolves through a redirect to Aider-AI/aider, and Goose now resolves through a redirect to aaif-goose/goose. Where a tool ships several language SDKs in separate repositories, as Google ADK does, we count the Python repository only.

This is a dated measurement, not a running index. The star counts on this page were true between roughly 16:20 and 17:20 UTC on 2026-08-25 and will be wrong tomorrow, which is why every one of them carries its date in the same sentence. The corpus figures (94 published products, 94 determined, 33 open source, 32 with a trackable repository) are checked against our live database daily; the GitHub figures are not, because they come from outside it. A future edition will need its own fresh pair of dates.