# 0 stars after a month of building, then 9 days after a launch post it reached 670: Crucible, a 3D engine for AI agents · GitHub open source project

> Source: https://stringers.io/en/github/gh_codepawl__crucible.html (Stringers / 什么值得做, generated 2026-11-20)
> Stringers tracks which small products suddenly took off each day, investigates why one by one, and distils playbooks from them. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these investigations and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools feed (daily breakouts) / plays (playbooks) / hot / products / data. Setup: https://github.com/komako-workshop/stringers (the user signs in with Google at stringers.io/en/start.html to get a personal URL). Site index: https://stringers.io/llms.txt

> The author posted a "Claude-made 3D work" every day on X for a month to build an audience, then posted one high-bookmark launch post on its own. Stars rose 224 the next day and it hit the Trendshift chart. Then the Claude Fable 5.2 launch came, and a 3D Zeno's paradox video rode the hype.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-15

codepawl/crucible · codepawl

# 0 stars after a month of building, then 9 days after a launch post it reached 670: Crucible, a 3D engine for AI agents

**What it is** An open-source JavaScript 3D engine with an MCP server. AI agents can build scenes, light them, preview them, and render out images or MP4 on their own.

**Why it took off** The author first spent a month posting a "Claude-made 3D piece" on X every day to build an audience. Then he posted one launch post with lots of bookmarks, which added 224 stars the next day and put it on the Trendshift chart. Claude Fable 5.2 launched around the same time, and a Zeno's paradox 3D film rode the wave.

**B** **How you could follow this** Don't build another 3D engine. Pick a domain where the finished output looks good and build a "tool for agents" there. Use it to make something every day for a month to build an audience, then post a launch post with a link on its own, timed to a new model release when demo slots are competitive.

Grade B: this playbook (ship output first, launch, ride a new model) worked for an author with only 2,500 followers, so anyone can copy it. But the 3D space is already taken by Blender MCP and vork, engines have a high bar, and the star count is only a few hundred. Whether it works still depends on whether the work is good enough and whether the timing catches a wave. Buzzing in niche circles

Data (updated daily) · repo created 2026-10-08 · 812 stars

GitHub · stars

Data 2026-11-15 → 2026-11-20 (6 days, 6 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

Crucible (formerly 67js) is an open-source JavaScript 3D engine with an MCP server. Agents like Claude Code and Codex can build scenes, light them, preview them, and output images or MP4. The author is an indie dev in Vietnam building it alone with agents. The first month after creation got zero stars. In Nov 8 a launch post went up, and 9 days later it had 670 stars. There was also a viral piece of about 24M views, yet it brought surprisingly few stars. The engine itself is hard to copy, but the playbook is learnable: use your own tool to make output every day, post a launch post on its own, then push again during a new model launch.

### A 31-second video

2026 Nov 8, Sunday, 17:23 (UTC), Vietnamese developer An Xuan (X @anxuanng) posted:

> Introducing 67js A 3D engine that any AI agent can drive, through 23 MCP tools (and counting) It's open source, so you can use it right away

with a 31-second video attached. He had only about 2,500 followers. The post got 74.6k views, 1,474 likes and 2,076 bookmarks, more bookmarks than likes. doomers.ai, a site that tracks launch posts, calculated 279 bookmarks per 10,000 views, 22.6 times the median of the 2,615 launch posts it tracks.

The repo gained 67 stars that day and 224 the next day (Nov 9), its biggest day so far, and it also reached #11 on the Trendshift JavaScript chart (daily star gains). The repo was created in Oct 8, and it was still at 0 stars in Nov 3.

### What it is

A full 3D engine written from scratch: real-time rasterization, path tracing, plus plugins for terrain, vegetation, characters, fluids, vehicles and video export. A scene is just data, with each object made of geometry, matrices and materials. Agents change the scene through MCP, get PNG previews back (up to four views side by side for comparison), and submit a render job once they're happy, for a still or an MP4. MCP doesn't accept arbitrary code or remote URLs.

The install instructions on the first screen of the README are written for agents, not humans:

> Let your agent do it. Paste this prompt into Claude Code, Codex, Cursor, Gemini CLI or OpenCode.

The prompt below tells an agent to install the package, read llms.txt and connect to MCP. There's also a command for the Claude Code plugin marketplace. The README says the goal is to be "as capable as Blender, as light as three.js," and includes a benchmark table: for the same scene, the bundle is 58.8KB gzipped, versus 133.8KB for three.js. Path tracing is compared against Blender's Cycles.

The whole repo has 3,804 commits and 947 merged PRs. Contributors are just the author and an automated dependency-upgrade bot (GitHub, 11-16). Outsiders have opened only 5 issues, from 3 people. So many people star it, but few actually use it. In Nov 14 it was renamed Crucible, and the old npm package @codepawlhq/67js was deprecated.

### The previous month: a new piece every day, and the engine still at 0 stars

Before the post went up, the author had been posting on X for a month. From Oct 9 on, almost every day brought a new "Claude-built 3D piece": a seal, a Jacquard loom, a playable data center, a Titanic cross-section, Arduino tutorials. Source code and prompts were public in a second repo, 3d-builds. Most single posts got a few hundred to a thousand views. The best one (the Oct 14 data center) got 11K.

The engine gained no stars that month, but he built up two things: a library of work, and an audience that enjoyed watching Claude make 3D.

In Nov 4 he switched hooks: he converted 1,033 real household items Google had scanned into 345MB GLB files and put them up for free on Hugging Face. The post got 52.9k views and 1,235 bookmarks. In the replies he linked the dataset first, then added "works in 67.js too" and the GitHub link. That day and the next, the engine gained 7 and 32 stars, its first real movement. He later posted twenty-odd more such packs, turning public but hard-to-use 3D datasets from Google, Amazon, Microsoft and NASA into formats that load directly. The most-downloaded one has 65k downloads (Hugging Face, 11-16).

### Where the stars came from

| Date | Stars gained | What happened that day |
|---|---|---|
| 10/8–11/3 | 0 | A new piece every day, and the engine didn't move |
| 11/4–11/7 | +52 | The free GLB dataset post |
| 11/8 | +67 | The "Introducing 67js" launch post |
| 11/9 | **+224** | #11 on the Trendshift JS chart; Claude Fable 5.2 released the same day |
| 11/10 | +91 | The Zeno's paradox 3D film, about 24M views |
| 11/11–11/13 | +179 | A repost from @RoundtableSpace (275k followers), and a batch of AI news sites wrote it up |
| 11/14–11/15 | +60 | Renamed to Crucible, posted 0.20.0 |

Data from Trendshift's daily star gains, about +674 in total. On GitHub it stands at 677 stars and 53 forks, with 11-16 as of the snapshot. The npm package had 6,950 downloads in the last month, peaking at 956 in Nov 10. That includes repeated installs from CI and agents, so don't read it as a user count.

Two other repos under the same org also grew: viz3d (a skill that has an agent turn a topic into a 3D piece, with Crucible as the default renderer) went from 25 stars in Nov 2 to 198 stars; and the 2D/3D skill collection creator-generalist went from 27 stars to 207 stars. The repos link to each other and share traffic.

### 24M views, for 91 stars

Nov 9, Anthropic released Claude Fable 5.2, and X filled up with "Fable 5.2 built X" demos. That same afternoon, a Turkish science account with 1.04M followers posted an old-school video: a father built an app to teach his daughter Zeno's paradox, showing why your finger can never actually reach the other finger. The post got 23.9M views.

19 hours later, on Nov 10 09:38 (UTC), An Xuan quote-posted it with his own 3D video: a zoom from the fingertip down to the atoms of keratin, atomic structures pulled from the Protein Data Bank, scale bars at true size. The caption was one line:

> Fable 5.2 built a 3D film to explain Zeno's paradox and show why your fingers never really touch

The post shows about 24M views, 9,973 likes and 4,311 bookmarks on X. Views nearly match the quoted original, which looks high for an account with 2,500 followers. Unverified. A community note later said it confused Zeno's paradox with atomic physics, and the controversy spread it further: over a dozen AI repost accounts on Instagram, Facebook and YouTube Shorts shared it, all crediting "@anxuanng on X"; the Turkish edition of Wired ran a story on Nov 12.

But the post body had no GitHub link. It only mentioned "rendered in 67.js" in a reply. The repo gained just 91 stars that day, less than half of the day before. Most people who watched the video remembered "the film Claude made" and not which engine was underneath. What actually sent people to GitHub was the launch post with 74k views.

### Who's competing for this spot

"Let an agent make 3D" is not a rare idea. Several repos have more stars than Crucible:

| Project | Launched | Stars | How it differs from Crucible |
|---|---|---|---|
| Blender MCP(ahujasid) | 2025-03 | 34.5k, +868 in the last 7 days | Lets an agent drive an already-installed Blender. Crucible installs via npm and renders in a browser or headless Chrome. |
| vork(voyager-so) | 2026-11-05 | 1,597 | A direct rival launched the same week: a native GPU engine written in Rust, selling cloud rendering. It has more than twice Crucible's stars. |
| three.js | 2010 | 118k | A general library for humans writing code. Lots of "Claude + three.js makes 3D films" tutorials use it directly. |
| MCPs for Unity / Godot / Unreal | 2025 | 2k to 16k | Drive existing game engine editors. Aimed at game developers. |
| manimgx(academa-labs) | 2026-10 | 2,800 | An animation engine for agents doing math and science explainers. |
| hyperframes(heygen) | 2026-03 | 84.7k | An agent tool that writes HTML and outputs video. The big winner in the 2D video lane. |
| ThreeBrowserStudio, chisel | 2026-06 to 08 | 26, 7 | Nearly the same idea. Nobody shared the output, so they never took off. |

Star counts come from GitHub, 11-16. The last row says the most: several repos are all "three.js scenes for agents," yet without anyone posting good-looking outputs they sit at single digits to a few dozen stars. The few hundred extra stars Crucible got came mostly from the author's content, not the product itself. The product's job was to make people who clicked in see a complete, serious engine, so they felt like starring it.

### If it were you

**One-line take:** Don't build yet another 3D engine for agents. Copy how it got distributed. Blender MCP and vork already hold this ground, and the technical bar for a 3D engine is high. But the playbook, "use your own tool to post outputs every day, batch up the works, then launch the repo separately and grab the demo slot right when a new model drops," worked for an author with only 2,500 followers, with no name and no big company behind him.

**Who it's for:** People already using Claude Code or Codex to build things, with a tool whose output you can see: charts, maps, animation, music, CAD, circuit diagrams, web effects. People who will post work on X every day for a month straight.

**Who it's not for:** Anyone whose tool's output can't be turned into a screenshot or short video. Pure backend or CLI libraries can't use this playbook. Anyone who wants real users and a community: stars from this approach are mostly passers-by who thought it looked cool. Crucible has 670 stars and only 5 external issues. Anyone who only wants 3D: this lane is already crowded.

**How to do it:**

1. Pick a domain agents still handle poorly but whose output looks great, and build an open-source tool that lets an agent drive it. Ideally the scene or data is in a format agents edit easily, and give the agent a way to check its result (screenshot, preview).

2. Once the tool works, don't launch yet. Use it to have the agent make one piece of work a day and publish it, with source and prompts in a separate public repo. Do this for a month.

3. Find data big companies publish but don't make usable, convert it to a usable format and release it free. Each release gets a post, with a tool link in the replies.

4. Once you have enough work and followers, post a separate "Introducing X": a video of about 30 seconds, one sentence saying it's for agents, and the link in the body.

5. Put agent-facing content on the first screen of the README: an install prompt you can paste straight to your agent, an llms.txt, a Claude Code plugin command. Add a benchmark table against established tools so experts trust it.

6. Change any gimmicky name early. 67js was renamed to Crucible when it had 600 stars, and the later you rename, the harder it gets.

**How to get users:** X is the main battlefield. On the day of a model launch and the day after, "X built Y" posts get amplified by launch buzz and repost accounts. Quote the hottest related post that day with a better version, and always put the repo link in the body. Crucible's 24M views only turned into 91 stars because that link was missing. After launch, get AI news accounts to repost it (like @RoundtableSpace), and submit the repo to Trendshift and MCP and skill directories. Crucible skipped Hacker News, Reddit and Product Hunt. You can cover those.

**When to do it:** Build the work in advance, but time the launch to the few days around the next major model release. "An X engine for agents" still feels new. Once someone in a niche gets a few thousand stars, followers can only compete on technical merit.

**Signs you're wrong:** If you post a piece every day for a month and single posts stay at a few hundred views, with nobody asking "how did you make this," your output isn't good enough, or nobody in the niche cares. Change niches before you launch. If the launch post's bookmarks are far below its likes, people think it's cool but don't plan to use it, so don't expect star counts to keep climbing. Conversely, if projects in the same niche that post no content but just ship features grow just as fast (like vork), then the topic's underlying demand is doing the work. Put your effort into features.

[GitHub repo ↗](https://github.com/codepawl/crucible)

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

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| ahujasid/mcp-for-blender | Open source | 2025-03 | 34.5k stars, +7 in the last 868 days | Lets an agent drive an already-installed Blender. Crucible ships its own engine, installs via npm, and renders in a browser or headless Chrome. |
| voyager-so/vork | Open source | 2026-11-05 | 1.6k stars, +7 in the last 500 days | Native GPU in Rust/Slang. The agent writes shaders as code. Sells cloud GPU rendering. Launched the same week and has more than twice Crucible's stars. |
| three.js | Open source | 2010 | 118.5k stars | A general library for humans writing code. Many "Claude + three.js makes 3D films" tutorials use it directly. Crucible pitches agent-friendliness and a smaller package. |
| Unity / Godot / Unreal MCP (CoplayDev/unity-mcp and others) | Open source | 2025 | 2k to 16k stars | Drives existing game engine editors. Aimed at game developers. |
| academa-labs/manimgx | Open source | 2026-10-05 | 2.8k stars | An animation engine for agents covering math and science explainers. Doesn't do photoreal 3D scenes. |
| heygen-com/hyperframes | Open source | 2026-03 | 84.7k stars | An agent tool that writes HTML and outputs video. Part of the 2D video lane. |
| instruktlabs/kiln | Open source | 2026-06 | 330 stars | Agents build single 3D assets procedurally. Doesn't cover whole scenes or the rendering pipeline. |
| SamG-Coder/ThreeBrowserStudio, EYamanS/chisel | Open source | 2026-06 to 08 | 26 stars, 7 stars | Nearly the same idea, but no content distribution, so it didn't take off. |
| MCPs for Meshy / Tripo / Scenario | Other | 2026 | Commercial products | Generative 3D models (diffusion), not an editable scene engine. |

Where users come from

- **Verified** **X launch post** : 11/8 posted "Introducing 67js": 74.6k views, 1,474 likes, 2,076 bookmarks (doomers.ai puts its bookmark rate at 22.6× the median for launch posts). The next day, +224.

- **Verified** **New-model launch wave + quote of a viral post** : Fable 5.2 11/9 launched. 11/10 quoted @dunyasalbilim's post (5.2M followers, 3M views) with a 1.04D Zeno film made in Fable 23.9. The quote shows about 24.1M views and nearly 10k likes. But the body had no repo link, and the repo gained only +91 that day.

- **Verified** **Founder posts a 3D piece on X every day** : from 10/9 on, nearly one a day (logged at codepawl/3d-builds). Single posts ran from a few hundred to 11k views. From these posts alone, the engine stood at 11/3 stars in 0.

- **Verified** **Free GLB asset pack (Hugging Face)** : 11/4 posted a Google scanned-objects pack that drew 52.9k views, with the engine link in the replies. On HF, 21 data packs; abo-glb has 65k downloads.

- **Verified** **IG / FB / YouTube repost accounts** : chatgptricks and more than a dozen accounts reposted the Zeno film crediting @anxuanng. The effect on GitHub traffic is not quantified.

- **Verified** **KOLs and AI news sites** : @RoundtableSpace (275k followers) posted 11/12, drawing 46k views. @cremieuxrecueil quoted it for 66k views. Wired TR, agentriot, karasukami and others covered it. Trendshift JS chart rank #11.

- Inferred **Agent-facing distribution (plugin, llms.txt, skill)** : The README's first screen is an install prompt for agents, with a Claude Code plugin marketplace entry. The viz3d skill uses it by default to render videos. Effect not quantified.

## codepawl/crucible (formerly 67js): how it got going

### Verdict

Crucible is an open-source JS 3D engine from CodePawl, a one-person studio in Vietnam (founder An Xuan, X account @anxuanng). It ships with an MCP server so agents like Claude Code, Codex and Cursor can build scenes, set lighting, check PNG previews, and output stills and MP4 themselves. Almost nobody starred it in the first month after creation (10/8 created, and by 11/3 it had only 0 to 1 stars). From 11/8 to 11/16, over 9 days, it went from about 0 to 668 stars, thanks to three things:

1. **11/8 posted a formal launch on X** : "Introducing 67js, A 3D engine that any AI agent can drive, through 23 MCP tools," with 74.6k views and a high bookmark rate. The next day (11/9) it gained +224, the repo's biggest single day so far, and it reached #11 on the Trendshift JavaScript chart.

2. **11/9 Claude Fable 5.2 launched, in the middle of a new-model demo season** : the next day he quote-posted a viral Turkish science account's post (23.9M views) with "Fable 5.2 built a 3D film…", attaching his own 3D film (Zeno's paradox: two fingers that never touch). The quote got about 24M views and nearly 10k likes. Afterward, a wave of AI repost accounts on Instagram, Facebook and YouTube credited "@anxuanng on X," and the Turkish edition of Wired ran a story.

3. **The differentiation is easy to explain** : "A 3D engine for agents, not for humans" fits in one sentence. The README's first screen lets you paste a prompt straight to your agent to install it. There's also a benchmark table against Blender and three.js. Together, these make people who read the post want to click through to GitHub and star it.

Still, the numbers were small. A viral post with nearly 24M views brought in only a few hundred stars, because the post body had no repo link and only mentioned "rendered in 67.js" in a reply. Stars mainly came from the 74k-view launch post, Trendshift's listing, and a few days of secondary spread (@RoundtableSpace with 270k followers and 46k views).

### Data card and cross-check

- Card: 2026-10-08,39 days since creation, 668 stars, 53 forks, MIT, daily peak +33 (our repo 11-16). The card history covers only 11-15 and 11-16, because the repo was renamed twice: anxuanng/67.js → codepawl/67.js → codepawl/crucible. Our repo didn't merge the old names' history.

- GitHub API(2026-11-16): 677 star, 53 fork; 3,804 commits, 947 merged PRs, 734 issues. The only contributors are anxuanng (3,793 commits) and dependabot. Only 5 issues were filed by outsiders (from 3 users), so real usage is thin, and development is almost entirely the founder working with an agent. GitHub's stargazers API (REST and GraphQL) can no longer return per-star timestamps, so the daily curve uses Trendshift's data.

- Trendshift daily star gains (https://trendshift.io/repositories/336165):

| Date | Stars gained | merged PRs | Notes |
|---|---|---|---|
| 11/1–11/3 | 0 | 25 / 30 / 118 | 0 stars in the first month after launch |
| 11/4 | +7 | 64 | Posted the HF scanned-object data pack on X the same day, 52.9k views |
| 11/5 | +32 | 21 |  |
| 11/6–11/7 | +6 / +7 |  |  |
| 11/8 | +67 | 76 | Launch post "Introducing 67js" on 17:23 UTC |
| 11/9 | **+224** | 20 | Ranked #11 on Trendshift's JS chart; Fable 5.2 launched that day |
| 11/10 | +91 | 86 | The Zeno's paradox 3D film, about 24M views |
| 11/11 | +78 | 37 |  |
| 11/12 | +45 | 96 | Reposted by @RoundtableSpace |
| 11/13 | +56 | 32 |  |
| 11/14 | +25 | 43 | Renamed to Crucible, posted 0.20.0 |
| 11/15 | +35 | 62 |  |
| Total | +674 | 747 |  |

- npm `@codepawlhq/67js` had 6,950 downloads in the past month: 11/8 peaked at 440,11/10 on 956; the new package `@codepawlhq/crucible` got 11/14 on its first day 472 (api.npmjs.org). These downloads include CI and agents reinstalling repeatedly, so they are not a user count.

- The same org's repos are also cross-promoting each other (ours): viz3d (agent skill, makes videos with Crucible)11-02 has 25 stars, reaching 11-16 with 196; creator-generalist (2D/3D skill collection)11-10 has 27 stars, with 11-16 reaching 207,11-11 single-day +80.

- On Hugging Face, codepawl has 21 free GLB data packs (HF API): abo-glb has 65k downloads, simready-glb 36k, pixel-atlas-glb 25k, scanned-objects-glb 17k. Each converts public 3D datasets from Google, Amazon, Microsoft, NASA and others into ready-to-load GLBs. Each release gets a post on X, with a reply that adds "works in 67.js too."

### Where users come from (growth levers)

1. **The founder posts daily on X** (confirmed). @anxuanng joined in February 2026, has 2,491 followers, and has posted 1,582 times. Starting 10/9, almost every day he posted a 3D piece built with Claude (a seal, a jacquard loom, a playable data center, a Titanic cross-section, Arduino tutorials…), with source and prompts published in codepawl/3d-builds. Most posts in this month got a few hundred to a few thousand views, and the best one got 1k (the 11 data center, 10/14). **These showcase posts did not bring stars to the engine:** by 11/3 the engine had only 0 stars. Their value was building a showcase library, demo material, and an audience that wanted to see Claude make 3D.

2. **Free 3D asset pack as the hook** (proven). 11/4 "1,033 real household items Google scanned, packed into 345 MB GLBs, free on HF": 52.9k views, 1,203 likes, 1,235 bookmarks. The reply led with the HF link, then said "works in 67.js too" and attached the GitHub link. The engine got +7 and +32 that day and the next. Later, "Microsoft's 115 rigged NPCs converted to GLB" from 11/14 drew 11k views. The play: borrow a big company's public data for credibility, give the assets away free, and the engine gets pushed along with them.

3. **A single launch post** (confirmed, and the most important). 11/8 (Sunday) at 17:23 UTC, "Introducing 67js": a 31-second video with 74.6k views (about 30× his follower count), 1,474 likes, 102 reposts, 79 replies, 2,076 bookmarks. doomers.ai data shows 279 bookmarks per 10k views, 22.6× the median across the 2,615 launch posts they track. This one post carried the +291 between 11/8 and 11/9.

4. **赶上新模型发布潮,蹭爆款** (已证实)。Anthropic 11/9 发布 Claude Fable 5.2(CNBC、TechCrunch),X 上随即出现大量「Fable 5.2 做了 XX」的 demo。11/9 15:00 UTC,土耳其科普号 @dunyasalbilim(104 万粉)发了一条老梗视频「一个父亲做了个 app 给女儿讲芝诺悖论」,2,394 万次浏览,4.55 万个赞。19 小时后(11/10 09:38 UTC),@anxuanng 引用转发这条,配上自己用 Fable 5.2 做的 3D 片(从指尖一路放大到角蛋白原子,原子结构取自蛋白质数据库,标尺是真实尺寸):约 2,413 万次浏览,9,973 个赞,736 个转发,4,311 个收藏(fxtwitter 接口读数;按他的粉丝量这个浏览数偏高,也可能和被引用帖的计数有关,未核实)。这条帖子被加了社区注释(说它把芝诺悖论和原子物理混为一谈),争议反而让传播更广:@cremieuxrecueil(34.6 万粉)引用后有 6.6 万次浏览;Instagram、Facebook、YouTube Shorts 上至少十几个 AI 搬运号(chatgptricks 等)转发,都注明「@anxuanng on X」;Wired 土耳其版 11/12 发了报道。**但帖子正文没放 GitHub 链接** ,只在一条回复里说「rendered in 67.js」(那条回复的浏览数没能读到),所以浏览转成 star 的比例极低,11/10 只涨了 +91。

5. **KOL and AI news site amplification** (confirmed). @RoundtableSpace (275k followers) posted "Someone just open sourced 67js…" with 46k views (11/12); @DanKornas (100k followers) posted it too, 1,352 views (11/15); other smaller accounts include @OzanTRAI and @balance0014. Between 11/8 and 11/15, agentriot, karasukami (Japanese), gyuha (Korean), hermes-ai, wpnews, shiplog, aionradar, Persian dadenevesht, and Vietnamese repo.tinix (#2 on Vietnam's chart) all wrote articles, most seemingly auto-generated from GitHub and X. Trendshift's JS chart at #11 added some exposure.

6. **Treating agents as a distribution channel** (confirmed, effect not quantified). The first screen of the README is "paste this prompt into Claude Code / Codex / Cursor / Gemini CLI / OpenCode." It ships llms.txt, the Claude Code plugin install `/plugin marketplace add codepawl/crucible`, and two SKILL.md files, and has been listed on claudepluginhub and agentskillshub. The viz3d skill defaults to Crucible for rendering, which funnels users to the engine.

7. **Not found:** HN (Algolia has no posts for crucible.codepawl.com or 67js; CodePawl's earlier Orglet got a Show HN with only 4 points), Reddit discussion, Product Hunt, paid promotion.

### Competitors

| Product | Where | Launched | Scale | How it differs from Crucible |
|---|---|---|---|---|
| ahujasid/mcp-for-blender(Blender MCP) | Open source | 2025-03 | 34.5k stars, +7 in the last 868 days | Lets an agent remote-control a Blender you already have installed. Crucible is a self-contained engine, installed via npm, rendering in the browser or headless Chrome with no Blender needed. |
| voyager-so/vork | Open source + cloud | 2026-11-05 | 1.6k stars, 7 days recent, +500,11-09 in one day, +282 | A direct competitor that launched the same week. Native GPU in Rust/Slang, agents write scenes and shaders as code, core API about 2,500 tokens, sells cloud GPU rendering. It has more than twice Crucible's stars. |
| three.js | Open source | 2010 | 118.5k stars | A library for hand-written code; many "Claude + three.js makes 3D video" tutorials use it directly. Crucible's README compares bundle size against it (58.8KB vs 133.8KB gzipped). |
| CoplayDev/unity-mcp, godot-mcp, unreal-mcp and others | Open source | 2025 | 2k to 16k stars | Remote-controls existing game engine editors, aimed at game development |
| pascalorg/editor | Open source | 2025-10 | 25.5k stars | A 3D building editor with MCP |
| tangency-labs/solidcraft | Open source | 2026-10-07 | 16.5k stars | Parametric CAD automated through MCP; a different use case from this one |
| academa-labs/manimgx | Open source | 2026-10-05 | 2.8k stars | An "animation engine for agents," focused on math and educational video |
| heygen-com/hyperframes | Open source | 2026-03 | 84.7k stars | "Write HTML, get video, built for agents": a big winner in the 2D video space |
| instruktlabs/kiln | Open source | 2026-06 | 330 stars | Agents build single 3D assets procedurally, with MCP and render review, not full scenes |
| iamlukethedev/Hermes3D | Open source | 2026-08 | 787 stars | A 3D engine based on Hermes Agents |
| SamG-Coder/ThreeBrowserStudio, EYamanS/chisel | Open source | 2026-06 to 08 | 26 stars, 7 stars | Nearly the same idea (an LLM-first three.js scene MCP; agents modeling with CSG), but no content push, and it never took off |
| MCPs for Meshy / Tripo / Scenario | Big tech / commercial | 2026 | Commercial products | Generative 3D models (diffusion), not an editable scene engine. |

**Summary:** The "agents do 3D" space is already crowded; MCPs that remote-control Blender and Unity have been at tens of thousands of stars for a while. Crucible's distinct feature is that it wrote a full engine from scratch (Glaze rasterization, Cook path tracing, plus plugins for terrain, vegetation, characters, liquids, physics, networking sync, etc.), scenes are JSON data, and the agent can see previews and submit MP4 renders directly. Similar small libraries (ThreeBrowserStudio, chisel) sit at single or double-digit star counts. The difference is not the technology, it's whether someone posts great-looking output every day.

### Why it's working

- **A person who posts content:** nearly every day for a month he posted a "Claude-made 3D piece," building up a showcase and an audience. Without content pushing it, the library sat at 0 stars.

- **A standalone launch post with high bookmarks:** the 11/8 launch post carried the biggest single-day gain (+224), which came a day before the Fable 5.2 hit.

- **Timing:** the week Fable 5.2 launched, "AI-made 3D video" was a hot topic on X and YouTube (Nick Saraev, Futurepedia and others all made videos). He inserted himself into the center of that topic the next day by quote-posting the viral post.

- **Free assets bring traffic:** converting big companies' public 3D data to GLB and posting it on HF, with one post per pack, so the engine was seen alongside the assets.

- **A positioning line that is easy to say:** "3D engine that AI agents drive," plus "as capable as Blender, as light as three.js" and a benchmark table, which made it credible to technical audiences.

### Differentiation, and whether it's the real cause

The differentiation is at the product level: this is a full engine designed for agents, not an MCP wrapper bolted onto existing software. Scenes are data, MCP does not accept arbitrary code or remote URLs, agents get previews and comparison grids, and can submit render jobs. There is also one-click plugin install, llms.txt, a minor version every week, and a benchmark table comparing Cycles and EEVEE.

But that is not the main reason it took off. Libraries with similar ideas got almost no attention, and the competitor vork gained stars faster the same week. What decided it was distribution: the founder's showcase content, free asset packs, the launch post, and the timing of Fable 5.2's release. The product's role was to make the claim "these videos were made with an open-source engine" hold up. People who clicked through saw what looked like a complete engine, and that is what made them willing to star it.

### What to borrow

- **Build the showcase first, then release the tool:** use the tool to produce shareable output every day, then, once viewers ask "how was this made?", publish a separate "Introducing X" launch post. The viral post should have included the link; it didn't, and 24M views went to waste.

- **Grab a demo slot on the day a new model launches:** titles like "<new model> built …" get amplified by traffic and reposter accounts after an Anthropic launch; quote the viral post the same day and offer a better version.

- **Turn public data into usable formats** : convert big companies' public but messy data (3ds Max, scanned PDFs) into standard formats, post them free on HF, write one post per dataset, and plug your own tools along the way.

- **Agent-first README** : the first screen says "paste this to your agent", with llms.txt and a Claude Code plugin marketplace, so agents do the install.

- **Rename early** : while you still have few users, swap a gimmicky name (67) for a serious one and set up redirects.

### Still unclear

- Whether the quote post's 24.1M views are accurate (fxtwitter endpoint; high for an account with ~2.5k followers), and how many clicks the reply line "rendered in 67.js" actually drove.

- Trendshift date basis: whether the +67 on 11/8 all comes from posts published after 17:23 UTC, or if it had traffic before that.

- Real usage: there are only 5 external issues, and npm downloads include repeat installs from CI and agents. Most stars likely came from people who thought the videos were cool, not from users.

- How much of the Zeno video and recent works actually used Crucible versus three.js. The work post says "rendered in 67.js", but many early works in 3d-builds are three.js apps.

- Whether vork takes the "3D engine for agents" slot (it has cloud GPU commercialization and already has more than 2 times Crucible's stars).

### Sources

- GitHub: [github.com/codepawl/crucible](https://github.com/codepawl/crucible), README; issue #1183 rename notice; GitHub API (commits, PRs, issues, contributor count, 2026-11-16)

- Trendshift:[trendshift.io/repositories/336165](https://trendshift.io/repositories/336165)

- Launch post: [x.com/anxuanng/status/2119480946340835771](https://x.com/anxuanng/status/2119480946340835771); doomers.ai analysis [doomers.ai/launches/codepawl](https://doomers.ai/launches/codepawl)

- Zeno video quote post: [x.com/anxuanng/status/2120088745139638375](https://x.com/anxuanng/status/2120088745139638375); original quoted post [x.com/dunyasalbilim/status/2119807224558260444](https://x.com/dunyasalbilim/status/2119807224558260444)

- HF scanned-object post: [x.com/anxuanng/status/2117859334608887928](https://x.com/anxuanng/status/2117859334608887928); NPC post [x.com/anxuanng/status/2121604068703342705](https://x.com/anxuanng/status/2121604068703342705)

- @RoundtableSpace:[x.com/RoundtableSpace/status/2120973657107579257](https://x.com/RoundtableSpace/status/2120973657107579257) ;@cremieuxrecueil:[x.com/cremieuxrecueil/status/2120350037654094116](https://x.com/cremieuxrecueil/status/2120350037654094116)

- Wired TR:[www.wired.com.tr/claude-la-yapilan-video-viral-oldu-parmakla…](https://www.wired.com.tr/claude-la-yapilan-video-viral-oldu-parmaklarimiz-aslinda-birbirine-hic-degmiyor-mu-h)

- Fable 5.2 launch: [techcrunch.com/2026/11/09/anthropic-releases-new-mythos-and-…](https://techcrunch.com/2026/11/09/anthropic-releases-new-mythos-and-fable-models-amid-rsi-concerns/)

- Works repo: [github.com/codepawl/3d-builds](https://github.com/codepawl/3d-builds); HF: [huggingface.co/codepawl](https://huggingface.co/codepawl)

- Competitor stars: our github repo_profile (2026-11-16); vork README

- npm: api.npmjs.org (@codepawlhq/67js, @codepawlhq/crucible)

- Secondary coverage: agentriot.com, karasukami.com, gyuha.com, hermes-ai.net, repo.tinix.ai

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