The official side didn't ship an Android version, so he built one the next day. How DSHA got 3000 stars among a dozen-plus peers, and how it was later overtaken

The day after DeepSeek open-sourced its agent framework DeepSeek Harness, a developer created a repo that packed a whole Ubuntu and dsh into a 279MB Android installer. 84 days later it had 3,053 stars and about 140k APK downloads. It is the most complete project in the "run dsh on your phone" track, but since Oct 30 it has no longer been the hottest one. Below: how it got started, why it was overtaken, and whether you can still win stars here.

The person who created the repo the second day

On Aug 2026 '13, DeepSeek open-sourced its official agent framework DeepSeek Harness (dsh), similar to Claude Code. It passed 95k stars in two days and reached 274k stars by early November. It was written for desktop Linux, with no official Android version. The v29 desktop release in Sep 0.2 only shipped for macOS and Windows.

On the afternoon of Aug 14, GitHub user qiannianhuanxiang created DSHA. The first sentence of the README states the problem it solves: dsh "was written for glibc Linux, and running it directly on Android hits a pile of issues: native modules won't compile, link(2) is blocked by SELinux, the sandbox won't start, and the frontend is laid out for desktop." His approach was blunt: bundle a complete Ubuntu 24.04 arm64, Node, and dsh, all offline, into the APK. Install it, tap once, enter your model key in the web page, and it works. No ROOT, no Termux, not a single command.

On Aug 22, he posted a video from his Bilibili account br2x titled "One-click install! DeepSeek's founder just released a powerful tool! Run it on your phone without root — DSHA, customizable with your own plugins," with 51.7k views (Bilibili, BV18F8U6aEbG). The title borrowed the DeepSeek founder's name, but this is a third-party project.

How much it actually grew

The curve has three phases: before early October, 20–30 stars per day; from Oct 14 it jumped to 80–120 per day; after Oct 29 it fell back to 45–60. The lift in the middle lines up with two events: on Sep 29 the official side shipped the v0.2 desktop version (reported by Sina and 36 36Kr), bringing in a wave of new users; DSHA followed on Oct 9 by bundling dsh 0.2. But no evidence was found for who directly drove the traffic (inference).

The most telling figure is downloads relative to stars: APK downloads are about 45 times the star count. Most open-source projects have many stars and few users; this one is the reverse, meaning people really installed it on their phones. The issues show who they are: someone reported "570 sessions can't be backed up" (#214,10-29), someone installed it on a budget vivo U3x with an older kernel (#206), and someone on Xiaomi HyperOS had their agent tap another app only to be blocked by a system window (#212). Most users aren't developers; they're students and players who want to use an agent on their phone.

The same thing, at least a dozen people building it

The Android port of dsh is one of the most-crowded opportunities right after it went open source. A tech blog post from Sep 6 (blog.mushroom.cv) surveyed four routes and noted that "the repos were created between 08-13 and 08-16" — the survey didn't even list DSHA yet, so it wasn't a top project in early September.

DSHA stood out from this pile not because of any exclusive feature. "No ROOT, no Termux" was being claimed by at least five peers at the same time, and woaiys3 had earlier made "let the agent operate the phone." It won on three things:

• Most complete. Others either required Termux knowledge, required a second computer, or offered a stripped-down environment. DSHA is a full glibc Ubuntu: apt works, native modules compile, and upstream plugins run without changes. The package is as large as 279MB, but for a user who just wants it to run, that's not a problem.
• Maintained like a product. 68 days, 28 releases; in late August it shipped almost one version a day. The README has a selling point line: "when something breaks, it tells you what broke." It offers component self-checks, on-demand repair, automatic backup, and a separate emergency environment, and when a plugin fails to start it names the culprit. Beginners who would otherwise give up when installation failed were caught by these features to a large degree (inference).
• A Chinese community. A Bilibili video, two QQ groups (in mid-October the README was changed to "Group 2 is full, please join Group 1"), Afdian posts offering "beta builds for those who donate to join the group," and a listing in the dsh community's launcher section. Almost no presence on X or in tech media; its audience is entirely inside the Chinese Android player community.

On how it sustains maintenance: the top of the README lists three sponsors, all AI model relay services, one with an invite link. Once users install dsh, their next step is to configure a model key, and relay services sell exactly that. The original author has already transferred the repo to the DSH-APP organization, and another contributor, ym2025szz, now maintains the releases.

On Oct 30, it was overtaken

dsh-Go was created on Oct 25 and went public beta on 30. The same day it took first place on Trendshift's all-language daily chart, gained 6k stars in 11 days, with a single-day peak of +2,198. It overlaps broadly with DSHA (it also runs the full environment locally), but adds remote-controlling a PC, sharing conversations, a plugin marketplace, and a desktop pet that writes code and does backflips. More importantly, its author is the creator of dsh-TUI (6,990 stars), a terminal plugin dsh officially recommended, so he already has a reputation. Zhidongxi and NetEase both wrote about it, and AI accounts on X reposted it.

DSHA's daily star growth began dropping right on that day, from around 100 to around 50. Its last formal release was in Oct 22; commits have continued since (11-05 still pushing), and the issue tracker holds 34 open problems, mostly backup migration and permission bugs in the new version.

This shows one thing: within a platform-spillover track, going first and being the most complete only buys you about two months of lead. The real ceiling is visibility, and visibility goes to whoever has a name when they enter the field.

If it were you

One judgment: there is something to learn here, but don't build "dsh on Android" again. This track was already crowded with a dozen-plus projects in August. Now there is a head with 11k stars, name recognition and press coverage, and another clone won't earn many stars.

What's worth learning is the playbook: when a desktop-only open-source hit appears, be first to ship a "mobile / no-environment one-click package" and catch the platform's spillover traffic. DSHA created its repo the second day; dsh-Go came two and a half months later yet overtook it on reputation. Both paths show the window for this kind of opportunity is measured in days.

Who it suits: people who can package runtime environments (proot, containers, Electron, WebView shells) and can ship a working version within 48 hours of a platform going viral; ideally with some voice of their own on domestic channels like Bilibili, QQ groups and Xiaohongshu.

Who it doesn't suit: people who want to polish slowly before releasing; and anyone with no name in the target platform's community. If a well-known author shows up in the track, you'll be pushed down like DSHA was.

How to do it: watch GitHub for projects open-sourced recently that gained tens of thousands of stars in a day or two but support only one environment. The next one might be an agent framework or desktop tool from another model company. Build the version for the platform it's missing, make it "install and tap once to use," don't worry about package size, and put self-checks, repair and backup in the first screen of the README as selling points. If you want to stay within the dsh ecosystem, the gaps in dsh-Go and DSHA show up in their issue trackers: backup migration, older device compatibility, and accessibility control on domestic systems (HyperOS, vivo). A small, focused tool or plugin for these has a better shot than yet another launcher.

How to get users: on the day you release, get into the platform's official community and plugin listing pages; post a short Bilibili video with a title that states it clearly: "No root, one-click install, run XX on your phone"; open a QQ group to collect feedback. When the group fills up, that's the best word-of-mouth screenshot.

When it's worth it: when you can ship a first installable version within 72 hours of a new platform going open source, and reach the fifth release within two weeks.

What would show you read it wrong: if DeepSeek officially shipped an Android version before 2027, this kind of third party would all fade out together. Conversely, if DSHA's daily star growth recovers to 80 or above within a month after dsh-Go took off, it would mean the "most complete, most stable" positioning still has its own audience beyond the head, and following the "more stable one" can also win stars.

---

Data and sources: GitHub repo pages and Releases (DSH-APP/DSHA, ccch1mneyyy/dsh-Go, etc., 11-06); DSHA README and issue tracker (11-06); Bilibili video BV18F8U6aEbG; blog.mushroom.cv (09-06); flowtivity.ai coverage of dsh's open source release; Afdian afdian.com/a/dsha_apk. Where it says "inference", it is inference.