AI coding agents clone and push to a single repository hundreds of times, only to get blocked by GitHub's rate limits — this frustration has been spreading across the developer community. The person betting on this concentrated load is Thomas Dohmke, who served as GitHub CEO until the end of 2025. On July 8, 2026, his company Entire Inc. released a preview of a decentralized Git network. It's a system that mirrors existing GitHub repositories to regional mirrors in a single step, offloading the read traffic generated by AI agents. Its benchmarks claim a proven figure of roughly 570,000 clones per hour against a single repository.
The 570,000-Clones-Per-Hour Benchmark: What's Behind the Numbers
The benchmark Entire published on July 8, 2026, involved a test in which 200 simulated clients spread across Frankfurt, Paris, London, and Dublin simultaneously performed shallow clones against a single repository. The result was roughly 570,000 clones per hour, which works out to 158 clones per second — one clone completing every 6.3 milliseconds. In the push test, Entire recorded 586 pushes per second, or about 2.1 million per hour when converted. In a mixed load test combining clones and pushes, throughput stabilized at roughly 470 operations per second. These figures have been independently reported with identical values by both SiliconAngle and DevOps.com, lending them a high degree of credibility.
Meanwhile, Entire has also drawn a comparison with "Origin," a new feature from rival Cursor. According to The Register's reporting, Entire's push throughput was 2.1 million per hour versus Cursor Origin's 81,000, and clone throughput was 570,000 versus 296,000. However, this comparison figure comes solely from The Register's reporting; the body of Entire's official blog at entire.io contains only a chart image, with no numerical figures stated in the text itself. A roughly 26x difference in push throughput and roughly 1.9x difference in clone throughput make for headline-friendly numbers, but as of now they cannot be confirmed against a verifiable primary source.
How AI Agents Started Breaking Git
Human developers clone or push only a handful to a few dozen times per day, but an AI coding agent repeats the cycle of cloning an entire repository, committing changes, and pushing every single time it executes a task. Even if a human developer performed the maximum of 50 operations a day, that would amount to only about 2 per hour when averaged out. By contrast, the roughly 570,000 clones per hour recorded in Entire's benchmark is the peak value from a load test in which 200 simulated clients hit the system simultaneously — a simple calculation puts it at roughly 280,000 times the human-equivalent scale. While this isn't the actual average access frequency of AI agents in real-world operation, it does substantiate the claim that in development environments where dozens or hundreds of agents like GitHub Copilot, Cursor, Codex, and Claude Code run in parallel, operations at this scale can concentrate within just a few minutes. Dohmke has said of this situation: "Centralized Git hosting became a bottleneck as agents scaled. The load of many agents and developers converging on a single server manifests as rate limits, high latency, and outages."
This is precisely the problem Entire's decentralized network is designed to solve. Development teams mirror their existing GitHub repositories to Entire in a single step. The canonical source of the code remains on GitHub, while AI agents clone and pull from whichever Entire mirror is geographically closest to the user.
The idea is to reduce simultaneous load on GitHub itself by absorbing the bulk of read traffic on the mirror side. Writes — that is, pushed changes — are ultimately synced back to the canonical GitHub repository, so from the development team's perspective, the repository's location appears to remain on GitHub as before. The preview is offered on a waitlist basis, and the operating regions are currently limited to just three: the US, the EU, and Australia. This, too, stems from the very design of physically distributing regional mirrors. The more mirror locations there are, the shorter the physical distance an AI agent needs to travel, and the lower the resulting latency.
Why the Man Who Left GitHub Is Now Targeting Its Stronghold
Dohmke served as GitHub CEO from 2021 and announced his departure on August 11, 2025. He then stayed on at GitHub through the end of 2025 to support the transition, and just half a year after leaving, on February 10, 2026, he announced the launch of Entire along with a $60 million seed round at a $300 million valuation. Converted to Japanese yen at the exchange rate as of July 8, 2026 (1 USD = ¥162.43), that's roughly ¥9.74 billion raised at a valuation of approximately ¥48.72 billion. It's considered one of the largest seed rounds ever in the developer tools space, with lead investor Felicis joined by well-known venture capital firms including Madrona, Basis Set Ventures, and 20VC.
That investor list also includes M12, Microsoft's internal venture arm. GitHub is a subsidiary of Microsoft, and here we have Microsoft's own investment arm putting money into a rival service — founded by GitHub's former CEO, who has publicly pointed out his own former company's infrastructure limitations. Madrona co-managing director Tim Porter has praised Entire as something that "could become the next great developer platform."
In this setup, the winners are Entire itself, its investors including Felicis, and the AI agent vendors who can now operate without worrying about GitHub's rate limits. Entire already supports integration with major agents such as Claude Code, Codex, Cursor, and Factory AI, giving agent developers an incentive to switch as well. Conversely, the ones losing out are likely GitHub itself and Cursor's Origin feature, which was singled out in the performance comparison. GitHub finds itself having its own limitations publicly called out by its former top executive, while Cursor has been confronted with a comparison showing it falling short on actual performance numbers.
Entire operates as a fully remote team of more than 40 people spread across nine countries, with plans to grow to around 60 by the end of the year. Its co-founders reportedly include Cole Driver, who served as GitHub's Acting Chief of Staff. It's a case of core GitHub personnel moving straight into the heart of a rival company — and the fact that a team deeply familiar with GitHub's internal workings is the one designing this decentralized infrastructure adds another layer of irony to the whole picture.
The Lessons of GitTorrent and Radicle, and the Detour Entire Chose
Git was designed in 2005 by Linus Torvalds, triggered by a licensing dispute with BitKeeper, as a distributed version control system that didn't presuppose centralized management. According to GeekWire's reporting, Entire cites a 2007 talk by Torvalds, framing its pitch as a return to the original principle that "Git is inherently distributed." Yet in practice, actual Git usage has long been consolidated around a single hosting service: GitHub.
Entire is not the first to attempt decentralized Git. GitTorrent, whose development began in 2015, aimed to apply BitTorrent's protocol to share repositories peer-to-peer, but was effectively abandoned by 2016. Radicle, which aims to decentralize Git using peer-to-peer architecture and cryptographic signatures, remains limited in scale as of April 2026, with public seed nodes handling roughly 8,000 repositories and 600 nodes per week, and has yet to achieve commercial traction. What both share in common is that they tried to completely sever developers from the existing network effects built around GitHub.
Entire's mirror-based approach differs on this point. Because it leaves the canonical source of the code on GitHub and offloads only the read traffic to regional mirrors, development teams can adopt it without changing their existing GitHub workflows or review practices. Rather than pursuing the "Git without GitHub" that GitTorrent and Radicle attempted, Entire has chosen "Git that works alongside GitHub." Whether this compromise can avoid the failure patterns of the past will depend on how many teams actually make the switch after the preview period.
When Will Japanese Development Teams Actually Be Able to Use It?
The preview's operating regions are currently limited to three: the US, the EU, and Australia. The pricing structure reported by GeekWire outlines only a broad framework — a free tier, an individual plan, and commercial plans combining seat-based and usage-based billing — with specific pricing and any timeline for expansion into the Japanese-language market still unannounced.
Entire has indicated plans to launch native hosting for both public and private repositories within the next few months, and says it ultimately aims for full decentralization and open-sourcing of the Git backend. Whether additional regions, including Japan, get added down this roadmap will likely depend on how the preview stage performs.
Entire's bet is logically sound in that it correctly reads the structural shift in which AI agents hammer Git at a frequency far exceeding that of humans. Technical details such as the sync consistency guarantees across regional mirrors or security assurances are not spelled out in any primary source. Whether Entire can actually breach GitHub's stronghold will be answered not so much by benchmark numbers as by a quieter but verifiable metric: how many preview participants actually move into production, and whether the available regions expand beyond just the US, EU, and Australia.