A product manager who hadn't written code in 20 years, and the company behind him, New H3C: how Pinvou got its 2,800 stars

The open-source desktop AI agent "Pinvou Agent" has been live for 105 days and has gathered 2,846 stars. It looks like another individual vibe-coding success story, but digging further, behind it is a team from New H3C (H3C) and a desktop AI supercomputer for sale. What drives stars for projects like this, what you can learn, and what you can't, we'll go through section by section below.

A forum post

In 2026 July 28, a user named hexin5898 posted on the developer community LINUX DO with the title "A product manager who hadn't written code in 20 years built an open-source desktop agent: Pinvou." He wrote that in 2026 Spring Festival he was ignited by OpenClaw ("little lobster"), then followed it to Claude Code: "From March to June, I sat in front of the computer fifteen or sixteen hours almost every day... I felt like I'd suddenly gone back to 20." In late March to early May he built two versions himself; the first "went wherever I thought of and quickly became a mess," so he started over. In early May he handed the project to a colleague named h3c-hexin, and "with the efforts of him and several other colleagues" the repo appeared on GitHub.

The name "Pinvou" was also one of the candidate names Claude Code generated in batch. The post's last line is a question: "If a desktop agent could really finish a piece of work for you, what would you most want it to finish?" The post got 26 likes, and the author followed up in replies all the way through mid-August.

This post carries LINUX DO's "open-source promotion" tag. Under community rules, the project must be fully open-sourced and the README must link to LINUX DO, and Pinvou's README does exactly that.

What it is

An AI workspace installed on your computer, written in Rust + Tauri 2 + React, under the MIT license, with installers for Windows, macOS, and Linux. The README's first line says it is "not just a chat window"; the end point of a task is a file, not a reply. It breaks down into three uses:

• Work: attach files, mount local knowledge bases, call MCP tools and skills, and hand back a reusable file;
• Design: generate posters and charts from a single sentence, then click elements directly in design mode to edit text, color, and size;
• Code: connect Codex, Claude Code, and Kimi via the ACP protocol to modify real projects.

The underlying engine is forked from a terminal coding agent written in Rust (CodeWhale, 40k+ stars). Pinvou is essentially a desktop shell on top of it, aimed at "ordinary non-developer users."

Who's behind it

The repo is organized under the name Pinvou alone, but the clues are easy to connect:

• The official site pinvou.com is titled "New H3C Intelligent Terminal · Complete Local AI Solution," and its homepage sells Pinvou alongside MegaCube. MegaCube is H3C's desktop AI supercomputer, with an NVIDIA GB10 chip and 128GB of unified memory, essentially H3C's version of the DGX Spark.
• H3C's official product navigation lists "Pinvou AI Agent Platform" directly, alongside MegaCube and MegaBook; the dealer pages sell "Pinvou × MegaCube."
• The top contributor asto18089 (416 commits) has a GitHub profile reading The New H3C Group, Guangzhou; another of the main contributors is named h3c-hexin.

So this is companion software for a hardware product. The features that other desktop agents don't really offer are aimed at MegaCube buyers: a built-in model store (download weights, deploy, check run status), GPU/memory monitoring, one-click connection to local vLLM; installers for LoongArch (64) chips from UOS and deepin; ready-made connectors for Feishu, DingTalk, WeCom, Tencent Meeting, Tencent Docs, and Qichacha; and model templates for DeepSeek, Kimi, Qwen, Doubao, and GLM. This is a configuration for government, enterprise, and Xinchuang (IT application innovation) customers.

Where the stars came from

The growth curve can be split into three phases:

Several things it has not had say just as much: it has never made GitHub trending, it can't be found on Hacker News or Product Hunt, no big accounts on X have shared it, and no media has covered it.

Two more numbers look off. First, the watch ratio: 302 watchers to 2,864 stars, about 10.5%, where typical open-source projects sit at 1–3%; the fork-to-star ratio of about 14.6% is also high. This looks like a group inside a company or among partners watching the repo (inference; GitHub's stargazers endpoint won't open for this repo, so it can't be verified). Second, downloads: over 30 releases are out on GitHub Releases, but every installer for each release combined has only 10–60 downloads. v0.10.17, released by 11-04, got 3 Windows downloads and 2 macOS downloads. Stars mean people looked and liked what they saw. They don't mean people are using it (most may be downloading from the website instead, unverified).

Who it's competing with

This category is already crowded. Big tech has Tencent's WorkBuddy, Alibaba's CoPaw, and the native desktop apps from Claude and Codex. Open source has a large group too:

Pinvou sits in the middle of the pack. The creator says so in their own post: "Turns out WorkBuddy already existed, and then Marvis showed up."

Can you follow it?

This is a GitHub project, and "follow" here means building an open-source project that can also earn stars. Breaking it down:

Hard to copy: a full-time team, a release cadence of two a week across 30+ versions, complete engineering with ADRs and SBOMs, all backed by a large company's budget. The users brought in by hardware channels and distributors are also out of reach for an individual.

Easy to copy:

1. The LINUX DO "open-source promotion" post. Almost zero cost: tag it, make it fully open source, add friendly links to the README, and tell a true story. "A 20-year product manager writes code for the first time" carries shareability on its own.
2. Write the English README so it can get reposted. A one-line positioning statement, one screenshot, a 90-second demo video, three language versions. The overseas reposts in August weren't solicited. The repost accounts found it themselves.
3. Claim the spots overseas products can't reach. Domestic office connectors, domestic model templates, and Xinchuang (government-procurement IT) installers are things Claude and Codex won't build, and they're the hardest parts of Pinvou to replace.

Not worth copying: building yet another "do-everything desktop agent workbench." There are dozens of these, four or five from big companies. An individual project struggles to stand out here. Pinvou, with a big company behind it, has only reached 2.8k stars and about twenty a day.

If it were you

Verdict: C. You can learn a cheap, effective open-source promotion playbook from it, but "general desktop agent" is already saturated, and about half of Pinvou's momentum comes from things you can't get (a corporate team and hardware channels).

Who it fits: individual developers who already have an open-source project and need the first batch of stars, especially Chinese-speaking developers; or small teams with hardware, compute, or cloud resources to sell who want to use an open-source tool as the entry point.

Who it doesn't fit: individuals trying to build a "desktop Claude Code / lobster shell" from scratch to collect stars. That path already has projects with 15k and 40k stars and a row of big-tech products.

How to do it: don't build a big, all-in-one workbench. Pick one narrow point that Pinvou touches but doesn't go deep on, and build it out fully. For example, "one-click deployment and monitoring of local models" (for DGX Spark, Mac Studio, and AI PC users), or "a collection of MCP connectors for domestic office software" (Feishu, DingTalk, WeCom, Tencent Docs). Narrow tools are easier to explain in one sentence and easier to get into awesome lists.

How to get users: first post on LINUX DO following the open-source promotion rules, and explain clearly why you built it. Prepare the README in Chinese and English with a screenshot and short video at the top. Then wait for overseas repost accounts and aggregators, and submit to lists like awesome-mcp and awesome-llm yourself.

Timing: the local AI hardware wave (GB10 all-in-one machines, AI PCs) is still rolling out, so building local model deployment tools is still early before the first half of 2027. Domestic office connectors can be built at any time, but big companies could add their own at any point.

Signs you've misread it: ① Pinvou suddenly accelerates over the next month or two, makes GitHub trending, and gains over a hundred stars a day. That would mean the "desktop agent workbench" still has room, and this analysis underrated it. ② Within a month, either narrow niche (local model deployment or domestic office connectors) gets a strong open-source project or an official big-company solution. That would mean the narrow angles are taken too.

Data and sources

• Daily star and fork numbers: GitHub, 2026-10-12 to 11-06. Watcher counts, Releases download numbers, fork timing distribution, and contributor profiles: check on GitHub, 11-06.
• LINUX DO post: linux.do/t/topic/2665306 (2026-07-28). Quotes come from the original post.
• New H3C relationship: the pinvou.com homepage and megacube.html, the h3c.com product navigation, and the MegaCube page on distributor yinzhongct.com.
• Overseas spread: X @Marco_Ramilli, LinkedIn "GitHub digest 14 August 2026", Instagram reel, YouTube Shorts.
• Competitor star counts: GitHub, around 11-06.