The GitHub League of Coding Agents: pi Passes 109,000 Stars While Gemini CLI Stalls

Six coding agents, one public counter: the GitHub stars. On 26 September 2026 at 07:03 CEST I pulled the numbers for the most important harnesses myself, straight from the GitHub API, eleven repositories in one pass, plus two searches. The result is not a leaderboard in the usual sense but a league running at three speeds. pi sits at 109,401 stars and has grown by almost twelve percent in four weeks. Google's Gemini CLI sits at 107,166 stars and has practically not moved since 10 September. OpenCode stands at 210,088, Codex CLI at 126,506, the Hermes Agent at 249,018. I am writing this number story not as a market analysis but as an operations report, because my cyberdeck runs exactly the three projects that sit in the fastest lane: Hermes as the conductor, pi as the worker, herdr as the docking station. Anyone working with agents can read from this league where the day-to-day is heading, often before any press report mentions it.

Before the numbers, the method, because star counts without provenance are worthless. I compared three points in time: the Pinggy list of open source CLI agents from 29 August, my own radar counts from 10, 23 and 25 September kept in the vault, and this morning's query. Stars measure attention, not usage and not code quality: a project can keep its stars while its development stands still, and an active project can stay under-observed. So I do not treat the absolute count as a verdict here; I treat the movement between dated measurement points as the signal. Two limits are on the table from the start. Three points are too few for a curve, so the reading stays an observation. And OpenCode has moved house: the address sst/opencode today points at the same repository as anomalyco/opencode, both queries return an identical 210,088 stars. I document the move so nobody double counts the number or mistakes it for two projects.

The table shows the GitHub API state, sorted by stars. The movement column compares against the oldest dated measurement points I have: the Pinggy list from 29 August and my own radar counts from 10 September. Where no earlier point exists, the table says so explicitly, because a movement nobody can date is not a movement.

These numbers read best as three lanes. The growth lane belongs to the lean harnesses: pi gains 11,401 stars in 28 days, herdr adds 9.6 percent in 16 days, and both deltas are cleanly dated because both starting points sit in the radar. The midfield lane shows Codex CLI at plus 6.3 percent and OpenCode at plus 4.0 percent since late August, solid movement for projects this size, with a notable detail: according to a radar note dated 25 September, OpenAI has taken Codex to release number 1,145, and the code was last pushed today. The stall lane is the actual news. The Gemini CLI arrived on 10 September with roughly 106,900 stars, still ahead of pi; sixteen days later pi passed it, and the Gemini CLI itself sits 266 stars above its old count. No collapse, but a drop from leader to bystander inside two weeks.

The OpenCode case belongs in any research training course. On 23 September I noted a contradiction in the radar file: a report from 15 September counted roughly 160,000 stars, the Pinggy list from 29 August counted roughly 202,000. One of the two numbers had to be wrong, and the entry stayed flagged, pending verification. The clarification takes a minute once you ask the primary source: the GitHub API returns 210,088 stars today, more than both older values, and the report was off by about 50,000. That is why I treat reports about star counts only as secondary sources with an expiry date: numbers wander, repositories relocate, and the only instance that counts today is the query made today. The lesson from my own radar reads: do not smooth contradictory secondary sources away and do not bin them; leave them openly flagged until an own measurement resolves them. That is exactly what happened here, with a date and a number.

Two voices from the scene fit these numbers. On 2 September David Ondrej published a video titled My Agentic Engineering Workflow, the state of his setup after 6,775 agent sessions by his own account. About pi he says it is the simplest, most minimal harness out there, the one OpenClaw is built on, and OpenClaw, at 390,531 stars, is the biggest project in the entire field. About the Hermes Agent he says it is the original among self-improving systems. What counts as a recommendation in his video is an operations status on my side: the Hermes Agent sits at 249,018 stars, second place in the league, and pi has not caught it, but with plus 11.9 percent it set the strongest movement among the large CLI agents. I keep that on record because it independently confirms my own tool choice for the third time, after the Ronacher conversation about the pi workflow and after my own stack report from early September.

For my own operation the table is not a fan chart but a risk overview. I run the three projects that grow strongest in this measurement: the Hermes Agent as conductor and long term memory, pi as the worker for coding tasks, herdr as the multiplexer that docks both. If the stall lane were the fate of all lean harnesses, pi would be the place to read it, and the opposite is happening. The Gemini CLI remains the example that a strong launch does not protect against stalling: three months after its start it was out front, sixteen days later it was overtaken, without the stars showing why. I draw no forecast from this, the series is too short and measures too little for that. What I draw is an observation rule for running my own setup: the same query, the same order, a fixed cutoff date, once a month, and any deviation above five percent gets checked before it feeds a decision.

The limits, explicitly. Stars count observers, not users: from stars you can read neither uptime nor install counts nor the number of merged pull requests, and aider shows the extreme case, close to 49,200 stars with four months without a code push. The OpenClaw figure comes from the repository search, not a direct repo call, so it carries the small caveat of a search mask. Three points are not a curve: anyone deriving a market claim from plus 11.9 percent in 28 days is skipping past the measurement limits. So this article makes no prediction for 2027, only a movement between two dated points, verifiable by anyone who queries the public repositories themselves. And a commitment belongs with that: next measurement, same procedure, otherwise the series is not comparable with itself.

October brings the repeat, same repositories, same order, a fixed cutoff date, plus for the first time a second source as a cross check against the Pinggy ranking. Three questions matter: does the stall lane stay a one-off at the Gemini CLI, or do more projects slide into saturation? Does Claude Code at 148,119 stars catch Codex CLI at 126,506, even though both come from the same house? And can pi shrink the gap to OpenCode, currently 100,687 stars, within a quarter? The next release cycles of pi and Codex will probably move the midfield lane again; since the model parade of 0.87.1 the changelog has been a reliable indicator. All numbers in this article come from public repositories, the query takes minutes, and the cutoff date stands in the first line of this section.

About the author: I am Marcel, a graphic designer running the HUNTER cyberdeck on a Google Pixel 6a. I pulled the numbers in this article myself from the public GitHub API on 26 Sep 2026 at 07:03 CEST; the older measurement points have been in the radar in the vault since 07 Sep 2026. Last reviewed on 26 Sep 2026.

Sources: GitHub API queries for earendil-works/pi, openai/codex, sst/opencode and anomalyco/opencode, NousResearch/hermes-agent, google-gemini/gemini-cli, anthropics/claude-code, All-Hands-AI/OpenHands, Aider-AI/aider, herdrdev/herdr, openclaw/openclaw and ryanalberts/best-of-agent-harnesses, each fetched on 26 Sep 2026, 07:03 CEST. Pinggy, Best Open Source CLI Coding Agents, star counts as of 29 Aug 2026, recorded in the radar entry of 07 Sep 2026. David Ondrej, My Agentic Engineering Workflow (after 6,775 sessions), YouTube, 02 Sep 2026, transcript read on 26 Sep 2026. Tech-insider report on OpenCode of 15 Sep 2026, cited in the text as an outdated secondary source and resolved against the primary source. Own radar counts of 10, 23 and 25 Sep 2026, vault, Themen-Radar.md.