# "The AI isn't bad at modeling, it just can't see": how a 647-follower UP host got his Blender plugin to 1,600 stars among a dozen rivals · GitHub open-source project

> Source: https://stringers.io/en/github/gh_sixtysevenlf__dsh-blender-plugin.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> 作者 10-15 发了一条讲「MCP 太慢、AI 看不见结果」的 B 站视频(现 9 万播放),10-16 发版把视口搬进侧栏,10-17 DeepSeek 新模型主打 3D 建模,10-18 进了 dsh 应用内的插件索引,想让 DeepSeek 建模的人手边就是它,一周涨了约 1,000 star。

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

sixtysevenlf/dsh-blender-plugin · sixtysevenlf

# "The AI isn't bad at modeling, it just can't see": how a 647-follower UP host got his Blender plugin to 1,600 stars

**What it is** A free open-source plugin for DeepSeek Harness that uses a direct channel to let an agent watch, click, edit and verify the Blender session you have open, with the viewport shown live in dsh's sidebar.

**为什么火了** 作者 10-15 发了一条讲「MCP 太慢、AI 看不见结果」的 B 站视频(现 9 万播放),10-16 发版把视口搬进侧栏,10-17 DeepSeek 新模型主打 3D 建模,10-18 进了 dsh 应用内的插件索引,想让 DeepSeek 建模的人手边就是它,一周涨了约 1,000 star。

**B** **How you could follow** Stop building a Blender plugin for dsh. Instead, pick a desktop app that AI can script but cannot see the results of (CAD, PCB, game engines, video editing), build deep integration with a visual, clickable interface inside a new agent platform, and pair it with a video that shows the pain in milliseconds and dollars. Ship just before or after a model launch and get into the in-app index.

Grade B: the play is clear and copyable without a big following, but the Blender slot in dsh is already taken, copycats and name-jackers have arrived, and its burst rode a single DeepSeek model launch. A follower would have to find software without a dominant player and hit a similar window, and would need to be good at making demo videos. Being talked about in small circles

Data (updated daily) · Repo created 2026-09-13 · 1,692 stars · 7 days +192

GitHub · stars

Data 2026-10-09 → 2026-11-08 (31 days, 30 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

There are a dozen or so Blender plugins built on DeepSeek Harness (dsh), and the earliest one launched a month before it. This one now has 71 stars. It picked up 1,643 stars in 54 days, which is 20 times the earlier one's count. It didn't invent the category. It just posted a video two days before DeepSeek released a new model and pushed "3D modeling", telling people: your DeepSeek has been able to model all along; the tooling was hiding it from you.

### A video two days ahead of the model launch

On the night of 2026 Oct 15, at 10:05 (Beijing time), Bilibili creator 「赛博雪豹」 posted a video with a very long title: 《Did you know your DeepSeek 4.1 Flash modeling ability has been online all along? MCP is what dragged it into 'can't model': direct connection 2.5 seconds to connect, 105 milliseconds to look, 6 paths……》 The account had only 647 followers at the time. Three days later the video had 37k views, 3,731 favorites and 303 comments (a roundup post from 什么值得买 10-18). As of 11-06, it had 90.8k views (per the Bilibili page).

The video's argument was simple: when you drive Blender with AI, the old way meant a round trip through MCP for every edit. The model could not see the result, so it edited blind, errors piled up, it looped, and the bill climbed. The fix was a direct TCP channel: grabbing a viewport frame takes 55–160 ms, and an "edit one step, look at it" cycle takes about 105 ms. At the end of the video the host open-sourced it as dsh-blender-plugin.

The next afternoon the same account posted an update video for v2.1, "DSH blender plugin update, making your deepseek model better" (11-06 views, 55.4k), moving Blender's viewport into dsh's sidebar. One day after that, on 17 Oct, DeepSeek released V4.1-Flash-1017. The official announcement said it was "optimized for DeepSeek Harness" and pitched capabilities from "making games" to "3D modeling," with the headline demo building the Forbidden City from a single sentence. That night, live tests of "one-sentence modeling" started flooding Bilibili.

The repo's author uses the GitHub account sixtysevenlf, with 19 followers, no bio and no X account. Neither side states whether the Bilibili account and the author are the same person. But the video says "I open-sourced it," and the series includes a video titled "My DSH Blender plugin," so it is safe to conclude they are the same person (an inference, high confidence).

### What it is

dsh is an agent framework DeepSeek open-sourced in Aug 13 with the slogan "everything is a plugin." The main repo now has 275k stars. dsh-blender-plugin lets the dsh agent operate the Blender session you already have open:

- **See** : a live mirror of the Blender viewport in dsh's right sidebar; the agent can grab a frame from any angle it wants.

- **Click** : click a part in the sidebar or in Blender and its object name gets inserted into your next message; the model can also ask you to click something.

- **Edit** : runs bpy in a persistent Python kernel, and errors can be rolled back.

- **Verify** : checks assemblies for interpenetration and floating parts, compares against reference images, and blocks export if verification fails.

- **Works without a GUI** : starts a headless Blender when there's no interface, and still produces renders.

15 model-side tools in total. It is also wired into dsh-TUI, dsh's second frontend: selected items become small tags in the conversation, and the footer shows whether Blender is connected and whether the AI is paused. Open source under BSD-3, free.

The install method is interesting: besides npm and the one-click install on the dsh desktop plugin page, the README also gives a prompt you can "copy to AI and let it install it for you". The only acceptance test is that `doctor` returns `kind=ok`: "Don't skip the check, and don't say it's installed until it has been checked."

### Where the stars came from

| Date | Event | Cumulative stars |
|---|---|---|
| 9-13 | Repo created | — |
| From ¥9-21 | A few dsh plugin lists came in over the following days | — |
| 10-04 | Submitted an inclusion PR to awesome-dsh-plugin | — |
| 10-15 | Bilibili video: "It's MCP that made it unable to model" | — |
| 10-16 | v2.1 release + update video | 231 |
| 10-17 | DeepSeek releases V4.1-Flash-1017, pitching 3D modeling | 338 |
| 10-18 | awesome-dsh-plugin merged the listing; the dsh-TUI author forked it | 429 |
| 10-19 | Shipped v2.2–v2.5 in one day; on X, @shantanugoel says it's better than Blender's official MCP | 548 |
| 10-20 | Single-day peak +232 | 780 |
| 10-23 | AlphaSignal wrote a short news item | About 1,100 |
| 11-06 | +187 in the last 7 days, +467 in the first 7 days, slowing down | 1,643 |

(Daily stars come from GitHub; there are no day-by-day numbers before 10-16. Events come from Bilibili, GitHub, and the original X post.)

The awesome-dsh-plugin list has 23k stars, and the index it generates is the one the Discover page in the dsh app reads. Once 10-18 merged the listing, users could search for Blender in dsh and install it in one click. How Discover ranks things isn't visible from outside.

It never made GitHub trending, never hit HN, got no major media coverage, and I found no sign of paid acquisition or star farming. Most of the forkers are real people with Chinese usernames, and their timing lines up with the videos and releases. In the English-speaking world there's just one X post and one AlphaSignal item.

### A dozen-plus players, so why this one

There are quite a few Blender plugins in dsh:

| Add-ons | Launched | Scale (around 11-06) | How it differs |
|---|---|---|---|
| CheshireJCat/blender | 8-14, the second day dsh went open source | 71 stars, 1,522 weekly npm downloads | A big bundle of skills plus helper scripts, with no live viewport or click-to-select |
| pearjelly/deep-blend | 9-13, same day | 102 star | Scene recipes and lighting for product shots, aimed at people who don't model |
| JOJO-Adam/dsh-blender | 10-19 | 0 star | Runs headless Blender only, with no UI linkage |
| xuezuoys/dsh-blender-mcp-bridge | 10-27 | Very small | Wraps an existing Blender MCP |
| blender-mcp (outside dsh) | 2025-03 | 33k stars | A general-purpose MCP that works with any model, with no UI |
| Blender's official MCP (outside dsh) | 2026 | Official | Requires Blender 5.1+, is awkward to install, and has no UI |

Interestingly, its Blender-side addon was adapted from blender-mcp, and the README's first instruction is to stop blender-mcp, because the two compete for the same 9876 port.

What pulled ahead came down to these things:

**It framed "why AI modeling doesn't work" as a story people could follow.** "It's not that it can't, it's that it can't see", backed up with millisecond figures. The line was later repeated word for word in science-explainer posts on What's Worth Buying and Zhihu, and became the standard explanation for this round of "AI modeling" talk.

**It did things MCP can't.** MCP only has tool calls, with no interface. dsh lets plugins take up the sidebar and the TUI footer, and it used that layer fully: you watch the AI model inside dsh and click to say "change this one". Other dsh plugins in the same niche are skill bundles or headless scripts; only this one puts the interface first.

**The timing lined up.** The video went up two days before the model launch. v2.1 shipped the day before, and the PR got merged the day after. The model makers did the job of teaching people that "AI can model 3D", so when users wanted to try it, this tool was already at hand. Whether that was luck or planning, you can't tell from the outside. The author 10-04 opened the PR, and the video title already said "4.1 flash", which suggests they were riding DeepSeek's release cadence.

**The author was the distribution.** 54 days, 421 commits, 30 releases, sometimes four versions in a single day. Each major version got a Bilibili video, and the same account also posted works built with the plugin: a Warhammer Dreadnought mech, a French villa, a 055 destroyer.

But one thing needs to be clear: **its star count is 20 times the earliest entrant's, while its usage leads by just over 2x.** The npm package only went up 11-01 later, and in its first 5 days it got 541–798 downloads a day. CheshireJCat's package does 1,522 a week. Even Blender plugins in the dsh ecosystem with only 0 stars get over 500 weekly downloads (a base load from mirrors and index crawlers), so download counts are only useful for comparing against each other. Everyone got the demand; the gap in stars mostly came from videos and visible demos.

### Others are already following

Oct 18 onward, a wave of copycats showed up: a HarmonyOS version dshm-blender (feature description nearly copied word for word), dsh-blender-cockpit, dsh-Blender, dsh-blender-helper, and someone ported its 15 live tools to another agent framework, pi. None have gained traction yet. Another name-squatter, dsh-plugin-blender-3d, had its npm version reported for hiding a crypto-mining script 10-26. Bad actors are piling in, which shows this category has real traffic on dsh, and also that it's already crowded.

### If it were you

**The takeaway:** "When a new agent platform opens up plugins, go deep-integrate a desktop tool people use every day into it, build an interface that's unique to the platform rather than yet another MCP, then post one video that explains the pain point clearly." The approach works, and it doesn't need a following (the author has 19 GitHub followers and 647 Bilibili fans). But the Blender slot in dsh is already taken, so building another dsh Blender plugin isn't worth it. What's worth doing is picking a different piece of software, or a different platform.

**Who it suits:** people who are heavy users of a specialized piece of software themselves and understand its scripting interface, e.g. developers who know Blender's bpy, Python for FreeCAD/KiCad, Godot, Unity, DaVinci Resolve, CapCut project files, and so on; people who can record their screen and explain a technical point in plain language; people who can handle shipping several versions a day.

**Who it doesn't suit:** people who only want to wrap an MCP and ship it. The thin wrappers in dsh are all in single-digit stars. People who want to "do better" in the dsh Blender slot: none of the late entrants since October has passed 100 stars. It's also a bad fit for anyone who can't make demos: in this category, stars follow the videos, not usage.

**How to do it:**

- Pick a piece of software that "AI can in theory handle but in practice is held back by the tooling". Test: the model can write its scripts, but can't see the result of execution. CAD, PCB design, game-engine scenes, and video-editing timelines all have this problem.

- Build the layer that's unique to the platform: bring the software's visual output into the agent's sidebar so users can click to select, and drop the selection straight into the conversation. MCP can't do this, and it's where same-category tools fail to pull ahead.

- Turn "fast" and "verifiable" into numbers: how many milliseconds to look at it, how many faces, how many dollars of tokens it cost. Those numbers become the video titles.

- Make installation dead simple: one click, plus a "let AI install it for you" prompt with a clear acceptance command.

- Doing the same thing on a different platform also works. In the plugin or extension systems of agents like Codex, Claude Code, and Cursor, it's worth checking whether the "Blender with a live UI linkage" slot is still empty (not verified).

**How to get users:** post a video on Bilibili (or YouTube) about the pain point, with the model name and specific numbers in the title; release an update video for every major version; post works made with your own plugin separately. Submit for inclusion to the index the in-app Discover page reads as early as you can. It matters more than any awesome list. Watch the model makers' release cadence, and ship versions and videos in the two or three days before and after a new model pushes a particular capability.

**When to move:** within the first two or three months after the platform opens plugins, ideally just before a model launch. dsh has been around for nearly three months, so what's left to build there is software without a leader. To switch platforms, start in the first month after that platform opens plugins.

**What would show you misread it:** ① your video passes 10,000 views but stars don't move. That would mean this wave came from the demand window of DeepSeek's model launch, not from the approach itself. ② Blender's official MCP or dsh's own team ships a UI linkage with a viewport. Then the "platform-unique interface" moat is taken back, and third parties are left with only deeper specialized features. ③ For the software you picked, AI still can't do it well even once it can see the result (for example, characters or organic forms; the comments ask "can it rig bones?" and the answer is "nothing can do that yet"). Then the problem isn't the eyes, it's the model, and connecting it faster won't help.

### Data and sources

GitHub repo page, README, releases, and daily star records (as of 11-06); npm download API (11-01 to 11-05); Bilibili BV13hJ26TEZK (published 10-15) and BV1ohJ86wEaQ (published 10-16), view counts checked as of 11-06; two roundup posts on What's Worth Buying, 10-18 and 10-19 (early views, favorites, comments, and the UP's follower count); DeepSeek API docs 10-17 and news announcements; X @shantanugoel 10-19; AlphaSignal 10-23; the awesome-dsh-plugin listing PR and the 10-28 security issue. Star and download counts for similar plugins come from GitHub and npm, as of 11-06.

[GitHub repo ↗](https://github.com/sixtysevenlf/dsh-blender-plugin)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| ahujasid/mcp-for-blender(blender-mcp) | Open source | 2025-03 | 33,472 star | A general-purpose MCP that works with any LLM and has no UI. This plugin's addon was renamed and adapted from it, and it requires users to disable it (the two compete for the 9876 port). |
| Blender Lab official MCP | Big-company features | 2026 | Made by Blender officially, and also Claude's Blender connector | Requires Blender 5.1+, hard to install, general-purpose with no UI; an X user called it 'very iffy' |
| CheshireJCat/blender(dsh-blender) | Open source | 2026-08-14 | 71 stars, 1,522 weekly npm downloads | Skill + helper bundle approach, with no live viewport or click-to-select. Entered a month earlier, but has 20 times fewer stars |
| pearjelly/deep-blend | Open source | 2026-09-13 | 102 stars, 525 weekly npm downloads | Product-scene recipes and light rigs, aimed at non-modelers |
| JOJO-Adam/dsh-blender | Open source | 2026-10-19 | 0 stars, 588 weekly npm downloads | Runs only headless bpy, with no GUI linkage |
| xuezuoys/dsh-blender-mcp-bridge | Open source | 2026-10-27 | 547 weekly npm downloads | Directly bridges the Blender MCP addon, a thin wrapper |
| Meshy meshy-dsh | Other | 2026 | Meshy official tutorial | Text-to-3D asset generation, does not drive Blender |
| arjun988/blender-skills / blenderwright | Open source | 2026 | 1,977 / 190 star | Skills or MCP for Claude Code / Cursor, not in the DSH ecosystem |

Where users come from

- Inference **Platform event: DeepSeek-V4.1-Flash-1017 released** : 10-17 The official announcement states it is optimized for DSH and pitches "from game creation to 3D modeling", with the headline case being a 3D model of the Forbidden City. Stars 10-16→10-20 rose from 231 to 780.

- **Verified** **Platform index / awesome lists (DSH Discover)** : 10-04 submitted a PR to awesome-dsh-plugin, 10-18 merged it. The index.json that list generates is the canonical index DSH Discover reads. 9-21 and 9-26 were also listed in two other lists.

- **Verified** **Self-made Bilibili videos** : 10-16 "DSH blender plugin update, making your deepseek model better" has roughly 55k views; the version videos for 10-21 and 10-30; the creator '赛博雪豹' started a "大肥鱼 modeling works" series on 10-19 (21.6k views).

- **Verified** **X developer repost** : @shantanugoel 10-19: 'looks quite good, compared to the official MCP', and said he would port it to pi.

- **Verified** **AI news roundup** : AlphaSignal covered it as 10-23, reposted by hermes-ai.net. 10-22→10-24 gained 223 stars

- **Verified** **Cross-plugin tie-in (dsh-TUI)** : the author of dsh-TUI (7,084 stars), 10-18, forked this repo, and 10-22 added a host plugin context chip API; this plugin hooked into it right after

- **Verified** **Long-tail SEO and content farms** : findthe.tools, skillsllm, yosvu.cn, CSDN gitblog, and YouTube Shorts index it passively

## Why sixtysevenlf/dsh-blender-plugin (DSH × Blender direct-connect plugin) took off

> Channel: GitHub. Research date 2026-11-06. Sources for numbers are in parentheses. "Our library" refers to the GitHub library and data cards on the agent-im machine.

### One line

它是 DeepSeek Harness(DSH)生态里的 Blender 插件:让 DSH 里的 agent 在你正开着的 Blender 里看视口、跑 bpy、自检、导出。它用的是自建 TCP 通道,不走 MCP。三件事叠在了 10 月 16–23 日这一周:DeepSeek 10-17 发 V4.1-Flash-1017,官方公告拿"3D 建模"当头号卖点;作者同一天前后在 B 站发了 v2.1 更新视频(5.5 万播放);插件 10-18 被收进 awesome-dsh-plugin 索引,DSH 应用内的 Discover 就是读这份索引。于是它成了"想让 DeepSeek 建模的人"默认装的那一个。它的 star 是同类最多那家的 20 多倍,但实际下载只领先 2 倍多。

### Key facts

- The repo was created 2026-09-13 by sixtysevenlf, a personal account registered 2026-03, with 19 followers, no X account, and no bio. There are 17 other repos, nearly all small DSH plugins (balance floating window, sound alerts, cost guard, divination chart matching, a Bilibili screening room, etc.), each with single-digit stars (GitHub API).

- Stars: 10-16 231 → 10-20 780 → 10-24 1,212 → 11-06 1,643 (our library). Single-day peak +232, on 10-20. Last 7 days +187, the prior 7 days +467, so growth is slowing. GitHub API 11-06 reads 1,663 stars, 65 forks. Has never made GitHub trending (trending_days_any=0).

- The fork timing matches the stars: only a few in September, then 2–6 per day from 10-16 onward (GitHub forks API).

- Shipping fast: 421 commits in 54 days, 30 releases (v0.4 → v3.0.6). 10-16 shipped v2.1.0,10-19 and then pushed v2.2–v2.5 in a single day (GitHub releases).

- npm: the package was first published 11-01, with 541–798 downloads per day from 11-01 to 11-05, 3,523 in total (api.npmjs.org). Note: in the same ecosystem, Blender plugins with very few stars still get 500+ downloads a week (mirrors and index crawlers add the base volume), so download counts can't be read directly as user counts.

### Competitors

| Name | Where | Launched | Scale | How it differs |
|---|---|---|---|---|
| ahujasid/mcp-for-blender (formerly blender-mcp) | Open source | 2025-03 | 33,472 star(GitHub API) | A general-purpose MCP that works with any LLM. This plugin's addon was forked from it (per the Bilibili v2.1 description: the addon was renamed from "MCP for Blender" to "DSH Blender Bridge", keeping the old name for compatibility). The README asks users to disable it, because both use the same 9876 port |
| Blender Lab's official MCP (blender.org/lab/mcp-server) | Big-company features | 2026 | Also a Blender connector for Claude; needs Blender 5.1+ and is hard to set up (see the comparison articles from mcp.film and strayspark) | Official, general-purpose, no UI. X user @shantanugoel called it "very iffy" while saying this plugin "looks quite good, compared to the official MCP" |
| CheshireJCat/blender (npm package dsh-blender) | Open source / DSH plugin | 2026-08-14 (a month earlier, the day after DSH went open source) | 71 stars, 1,522 weekly npm downloads | Takes a skill + helper route (30 skills / 13 runtime tools), with no live viewport and no point-and-select |
| pearjelly/deep-blend | Open source / DSH plugin | 2026-09-13 (same day) | 102 stars, 525 weekly npm downloads | Leans toward product-scenario recipes and a showcase gallery, aimed at non-modelers |
| JOJO-Adam/dsh-blender | Open source / DSH plugin | 2026-10-19 | 0 stars, 588 weekly npm downloads | Runs only headless bpy, with no GUI linkage |
| xuezuoys/dsh-blender-mcp-bridge | Open source / DSH plugin | Included in 2026-10-27 | 547 weekly npm downloads | Bridges directly to the Blender MCP addon |
| walkzzz/dsh-plugin-blender-3d | Open source / DSH plugin | 2026-08 | Small | 10-26's npm 0.2.1 was reported for a postinstall that mines XMRig (awesome-dsh-plugin issue, 10-28) |
| Meshy's meshy-dsh | DSH plugin / commercial | 2026 | Meshy official tutorial | Text-to-3D model; does not drive Blender |
| arjun988/blender-skills, HoldMyBeer-gg/blenderwright, etc. | Open source | 2026 | 1,977 / 190 star | Skills or MCPs for Claude Code / Cursor, outside the DSH ecosystem |
| lhmd/dsh-director-toolkit, Dogwind221/DWcinematic-whitebox | Open source | 2026-08/09 | 9 / 6 star | Produces directing plans and whitebox previsualization; does not connect directly to Blender |

Verdict: a dozen or so players do Blender inside DSH. The earliest, CheshireJCat, came a month before this one, but its stars are stuck in the double digits. In the general agent ecosystem, the same job is done by blender-mcp (33k stars) and Blender's official MCP. It didn't invent the category; it wins through a single-point bet on DSH, a new platform.

### Growth tactics (chronological)

1. **DSH platform tailwind (background)** : DeepSeek open-sourced DSH (deepseek-ai/deepseek-harness, 275k stars) on 2026-08-13, with the "Everything is a Plugin" design. Plugin ecosystem exploded from August. 27,736 repos are tagged dsh-plugin (GitHub search), and Zhihu, CSDN and Bilibili have a batch of "must-install N DSH plugins" lists.

2. **Awesome lists and index sites** : 9-21 is listed in Awesome DeepSeek Harness; 9-26 was picked up by Dominic789654's automated scan; 9-30 submitted a PR to hashgraph-online/awesome-ai-plugins on its own; 10-04 submitted a PR to awesome-dsh-plugin (23,280 stars), **10-18 merged** . The index.json that list generates (awesome-dsh-plugin.com/index.json, 4,703 plugins) is the canonical index DSH's in-app Discover reads (see the security issue on that repo, 10-28). After the merge, users searching for Blender in the DSH app can install it in one click. The author also built three awesome-list repos of their own (awesome-dsh-plugin, Awesome-DeepSeek-Harness-Plugins, a fork of awesome-ai-plugins), each with 0 stars, which did little.

3. **Bilibili self-made videos (by the author or accounts linked to them)** : On 10-16, a video titled "DSH blender plugin update, making your DeepSeek better at modeling" (v2.1, the page shows 55K views; Jina crawled m.bilibili.com). The same collection also has a v2.6.0 update (10-21) and v2.6.2 (10-30), plus titles like "Did you know your DeepSeek 4.1 flash modeling ability was always online? It's MCP that made it 'unable to model'". Bilibili user "Cyber Snow Leopard" 10-19 started a "Big Fat Fish modeling works" series: using this plugin plus V4.1 Flash 1017 to model Warhammer Dreadnought, Mechanicus Priest, and a French villa (21.6K views). One video in the series is titled "My DSH Blender plugin", so this account is very likely the author's own, unverified.

4. **Platform event: DeepSeek-V4.1-Flash-1017 (10-17)** : The official announcement (api-docs.deepseek.com/news/news261017) states it is "optimized for DeepSeek Harness," with end-to-end project capability "from game creation to 3D modeling," and the headline case is "3D modeling of the Forbidden City." The announcement doesn't mention this plugin, but it got many DSH users to try "having DeepSeek model," and modeling means connecting to Blender. Stars rose from 10-16 to 10-20 (231 → 780), right within that window.

5. **A repost in the English-speaking X crowd** : @shantanugoel posted 10-19 praising it as better than the official MCP, and said they'd port it to pi (x.com/shantanugoel/status/2112227456635871559). Follower and engagement counts are unverified. Our x-radar corpus turns up no other posts about this project, so there's little discussion on X.

6. **AI news roundup** : AlphaSignal covered "DeepSeek Harness Lets AI Build Full 3D Scenes Autonomously in Blender" on 10-23, reposted by hermes-ai.net. From 10-22 to 10-24 it gained 223 stars.

7. **Cross-plugin tie-in** : the author of dsh-TUI (7,084 stars, the officially recommended TUI), ccch1mneyyy, 10-18 forked this repo, and 10-22 added to dsh-TUI an interface letting host plugins drop context chips into the conversation (dsh-TUI PR). This plugin then built selected-chip, footer status, @ mentions, and a /blender command for the TUI, giving it exposure in the second-largest frontend.

8. **Long-tail SEO and content farms** : findthe.tools, skillsllm.com, yosvu.cn, and CSDN gitblog carry "let AI install dsh-blender-plugin for you" pieces (59 reads), plus a small YouTube Shorts video, "AI Controls Blender in Real-Time?!" (10-29). All are passive indexing.

No evidence found of paid ads, star-buying, or influencer outreach. Most forkers are real people with Chinese-language usernames, and fork timing lines up with the videos and releases.

### Why this one (the reasons)

- **Claim the "DeepSeek can model" entry point before anyone else** : 10-17 officially pitched 3D modeling as a selling point, so users needed a Blender channel that works right out of the box. It had just shipped v2.1 (sidebar viewport, select parts to write into the next prompt, multi-select) 10-16, and it got into the Discover index 10-18.

- **A form factor MCP can't match: watching the AI model inside the DSH interface** : The live Blender viewport mirrors into DSH's right sidebar; clicking a part inserts its object name into the message, and the model can also ask you to click something (blender_rt_pick). Whatever you select in Blender goes straight into the conversation. MCP only has tool calls, no interface. DSH's "everything is a plugin" lets plugins take the sidebar and the TUI footer, and this plugin fills that layer. "Not MCP, a direct connection" became its main public slogan.

- **Technical depth and a "verifiable" narrative** : 15 model tools and a 34-family / 228-op contract layer, with a headless worker, transactional rollback, assembly-clipping audits, and a render-acceptance gate that blocks delivery when it fails. The videos show measured millisecond timings and spec_hash, "no staged shots." This engineer-speak suits the GitHub crowd and led people on X to compare it with the official MCP and conclude it's better.

- **The author provides distribution** : The author makes Bilibili videos, submits PRs to lists, maintains their own lists, and ships releases almost daily. Each release earns another spot in the "recently updated" feeds and gives the videos new material.

- **Installability** : Offers an install path where you paste a prompt into your agent and let the AI install it for you. The only acceptance test is doctor returning kind=ok. Also ships a one-command npm install and a DSH Desktop plugin page, so the barrier is very low.

### Differentiation

Compared with generic Blender MCPs (blender-mcp, Blender's official MCP), it's tied to DSH, which is what lets it do UI: sidebar viewport, click-to-select, TUI chips, plus an acceptance gate and a headless fallback. Compared with the other Blender plugins in DSH (CheshireJCat, deep-blend, JOJO-Adam), it's the only one that's GUI-first and live-linked. The others are skill packs or headless scripts.

The differentiation is why it took off, but only half of it. The other half is timing and distribution: DeepSeek's model launch taught people that "AI can model," and DSH's Discover index made it installable in one click. CheshireJCat entered a month earlier and still gets 1,522 weekly downloads, so demand isn't the problem. The gap is mainly stars: it has 1,600+, while its rival has only 71. The star gap is much bigger than the usage gap, and it comes from the author's content distribution and a visible demo.

### How many people have seen it

Word of mouth is small: the Chinese DSH community has a Bilibili video that peaked at 55k views; the English side has one developer's repost on X and one AlphaSignal newsletter mention. Not on HN, not on GitHub trending, no major media coverage. Open source and free, so no revenue.

### Copycats

- Ports: stollor/pi-blender-rt (10-04), which brings its 15 real-time tools over to pi. @shantanugoel also said they'd port it to pi.

- Clones / same-period: JOJO-Adam/dsh-blender (10-19), chenchen-008/dshm-blender (10-18, a HarmonyOS version whose feature list nearly copies it: view the viewport, read the scene, run Python, screenshot verification), xuezuoys/dsh-blender-mcp-bridge (10-27), ToEntropyProduction/dsh-blender-cockpit (10-22), Century-Ula/dsh-Blender (10-22), GA17-LHY/dsh-blender-helper (9-26).

- Secondary forks: FuRongJun-1999's lingshu "world model bridge," and dsh-blender-wm-bridge, which the author forked for their own submission.

- Name-squatting: a version of walkzzz/dsh-plugin-blender-3d on npm was reported for hiding a crypto-mining script. It shows the Blender category already has enough traffic in DSH to attract bad actors.

### What others can borrow

1. When a new agent platform opens up, go after deep integration with a high-value desktop app (Blender / CAD / Unity / CapCut, etc.), and build a UI form unique to that platform instead of yet another MCP.

2. Watch model vendors' launch announcements for the capability they push (this time, 3D modeling), and have a plugin ready to carry that capability in advance. Ship the release, the video, and the index listing within 48 hours before or after launch.

3. Getting into the in-app Discover index of the platform matters far more than star counts on an awesome-list.

4. Sell it on verifiability: real millisecond timings, hashes, doctor kind=ok. Add a "let AI install it for you" prompt, and the install barrier drops.

5. Ship often, and pair each release with a set of short videos. Every version update is another round of distribution.

### Unknowns

- The day-by-day stars before 10-16 can't be recovered: the stargazers API returns 404 for every repo, and our records only start at 10-16. So whether there was a small bump from September to early October is unverified.

- Whether the Bilibili uploader is the author, and what "赛博雪豹" is to the author.

- View counts on the page were scraped by Jina and may be inaccurate. @shantanugoel's post engagement and follower count are unverified.

- Unclear how much of the npm downloads come from mirrors, index crawls, and frequent version releases. Comparing weekly downloads with CheshireJCat's 1,522, the real usage lead is about 2x, not 20x.

- DSH Discover's ranking rules (whether it sorts by stars) are unverified.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.