# Polyfork

Aiolum story: https://aiolum.com/story/polyfork/ · added 2026-10-25 · updated 2026-10-25

**In short.** Polyfork (Library of low-poly 3D assets written as code by AI agents, with an MCP server, Germany): US$16.3K in verified revenue in under 3 months, October 2026. Evidence grade A: disclosed by a company or a filing, or verified by a payment processor. Team: 1 founder; TrustMRR lists a team of 1, and AI agents build the assets. Launched 1 August 2026; still running, revenue falling.

Library of low-poly 3D assets written as code by AI agents, with an MCP server · Germany · AI-native · solo founder · still running, revenue falling · group: Build software with AI

- Outcome: US$16.3K in verified revenue in under 3 months, October 2026 (evidence grade of the headline number: A; headline in US dollars: 16307)
- Launch: 1 August 2026; months from launch to the outcome: 3
- Team: 1 founder; TrustMRR lists a team of 1, and AI agents build the assets
- Outside funding: No investors found; US$500,000 XPRIZE prize money in September 2026
- Main factor: Timing
- Meters, 1 to 4: evidence strength 3, copyable pattern 2, capital needed 1
- Tags: AI-native, Solo founder, Bootstrapped, Running, B2C, 3D assets

**The product.** A web store of 2,634 low-poly 3D models for games, websites and virtual reality. Each model is a small JavaScript program with settings (colours, sizes, number of parts) that rebuilds the model when a setting changes, and it also downloads as an ordinary GLB, FBX, USDZ or OBJ file. An MCP server lets AI coding tools such as Claude Code, Cursor or Codex search the catalogue and place models in a scene. A factory of AI agents writes and checks the models: XPRIZE wrote in September 2026 that it publishes more than 175 a week.

**Why it worked.** The founder made the 3D models themselves into code, so AI agents could write them in large numbers and other AI agents could read and change them. He opened the store with a one-time Founders Club seat for US$100, and Stripe data shows US$8,917 in the first month. Revenue then fell each month, and almost all of it came from one-time seats, not from subscriptions.

Evidence grades: A = disclosed by a public company, a filing or a payment processor, or verified by one; B = two or more independent reports agree, or a journalist confirmed it; C = a founder claim that no one else verified; D = reported, not verified: one founder claim or a third-party estimate with no second source; calc = computed on this page from graded facts.

## Summary

Polyfork sells low-poly 3D assets: 3D models of buildings, cars, characters, trees and terrain built from few triangles, so that they load fast in a web browser or a game. Each asset is a small ES module, a JavaScript file that another program can import, and calling it returns a ready model for three.js, the most used library for 3D graphics on the web. The module declares typed settings ("knobs") such as colour scheme, height or number of parts, and the same model is also sold as a GLB file for Godot, Blender and other engines that read glTF, the standard 3D file format, and as FBX for Unity and Unreal. For AI coding tools there is an MCP server: MCP, the Model Context Protocol, is a standard way for an AI assistant to call outside tools, here search_assets, get_kit, buy_plan and nine more.

Lucas Martinic, an XR developer who represented Germany in the Build with Gemini XPRIZE, started the store on 1 August 2026, by the date on his TrustMRR profile. The models are not made by hand: XPRIZE wrote that "an AI-agent factory publishes more than 175 a week", and the company's own posts describe a "builder-and-judge lane" in which one AI agent writes a model and automatic checks and a judge test it. The catalogue grew from 326 assets on 5 August to 2,634 on 25 October 2026. On 26 September 2026 Polyfork won the first prize of the Build with Gemini XPRIZE, US$500,000.

The headline number is grade A. TrustMRR reads it from Polyfork's Stripe account: US$16,307 from August to 25 October 2026, made of US$8,917 in August, US$5,780 in September and US$1,610 in October to the 25th. The trend is down, and the revenue is almost all one-time payments: TrustMRR shows US$0 of monthly recurring revenue and no active subscription, and the seat counter on the pricing page shows 145 of the US$100 lifetime seats taken. The page also shows four failed companies that sold 3D content creation or 3D tools for the web.

## Pricing

Polyfork plans on 25 October 2026, read from the company's pricing page (grade C)

| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Free | 0 | 0 | every asset under 600 triangles, 883 of 2,634 |
| Pro | 12.42 | 15 | every asset, own CDN path, remix baking |
| Lifetime |  | 100 one-time | everything in Pro for life, 200 seats |
| Platform 400 | 249.17 | 299 | 400 assets, up to 5,000 of your users a month |
| Platform 1000 | 582.50 | 699 | 1,000 assets, up to 25,000 users a month |
| Platform 2000 | 1,249.17 | 1,499 | 2,000 assets, up to 50,000 users a month |
| Full catalogue |  | from 2,900 | priced on a call, up to 100,000 users a month |

The pricing page shows Pro at US$15 a month and says yearly billing saves US$31; the yearly column is calculated from that (15 times 12 minus 31 equals US$149 a year, about US$12.42 a month). Pro was US$10 a month in August 2026. The Lifetime seat is one payment of US$100 (shown as reduced from US$125), capped at 200 seats including the 100 sold-out Founders Club seats, with 55 left on 25 October 2026. Students and people in low-income and lower-middle-income countries can apply for Pro at US$50 for 2 years. The Platform plans are for products that hand the assets to their own users; their yearly prices are ten months for twelve. Separate orders: a custom character skin costs US$9, a custom kit of 60 parts US$85 if it is published to the catalogue, or US$200 if it stays private.

## How much revenue does Polyfork make? The numbers, graded

US$16,307 of Stripe-verified revenue between 1 August and 25 October 2026, falling each month, almost all from one-time US$100 seats

| Fact | Value | Grade | Note | Sources |
|---|---|---|---|---|
| Verified revenue, 1 August to 25 October 2026 | US$16,307 | A | Stripe data read by TrustMRR, last synced at 17:13 UTC on 25 October 2026. The same figure is shown for all time, the last 3, 6 and 12 months, so no revenue was recorded before August 2026. | [1] |
| Revenue in August 2026 | US$8,917 | A | The first month. About US$288 a day (calculated: 8,917 divided by 31). | [1] |
| Revenue in September 2026 | US$5,780 | A | 35.2% below August (calculated). | [1] |
| Revenue in October 2026, to the 25th | US$1,610 | A | About US$64 a day, 66.6% below September's daily average of about US$193 (calculated: 1,610 divided by 25, against 5,780 divided by 30). 8 of the 25 days had no revenue. | [1] |
| Revenue in the last 30 days | US$3,780 | A | As shown on the profile. The 30 daily rows listed, 26 September to 25 October, add up to US$3,580 (calculated); a separate stored snapshot shows US$4,780. Either way it is above US$2,000 a month. | [1] |
| Monthly recurring revenue and active subscriptions | US$0 and 0 | A | TrustMRR snapshot on 25 October 2026. The revenue is one-time payments, although a monthly Pro plan is on sale. | [1] |
| Daily payments in October 2026 | mostly US$100 to US$125, or US$10 to US$50 | A | 11 of the 25 days show US$100 or more. These sizes match one US$100 lifetime seat, plus small orders such as a US$9 skin or a US$15 Pro month. | [1, 3] |
| Lifetime and Founders Club seats taken | 15 on 5 August, 31 on 12 August, 136 on 3 October, 145 on 25 October 2026 | C | Read from the seat counter on archived and live copies of the home and pricing pages (100 seats at first, 200 later, counted as 200 minus the seats left). The counter is the company's own display. | [13, 14, 15, 3] |
| Share of revenue from US$100 seats | about 89%, at most | calc | Calculated: 145 seats times US$100 equals US$14,500, which is 88.9% of US$16,307. It is an upper estimate: some seats may have been paid at a lower price after a Pro credit or a reduced-price application. If the 200 Lifetime seats were in addition to the 100 Founders Club seats, the seats alone would be US$24,500, more than all revenue, so the counter must include them. | [1, 3] |
| First-month figures given to XPRIZE | US$4,428 from 54 paying customers, with 583 models | C | XPRIZE announcement, 26 September 2026. The window of that first month is not stated; it does not match the Stripe figure for August (US$8,917). About US$82 per customer (calculated). | [4, 5] |
| Assets in the catalogue | 326 on 5 August, 457 on 12 August, 1,921 on 3 October, 2,634 on 25 October 2026 | C | Read from archived and live copies of the home page. The live page shows 256 new assets in the last 7 days. | [13, 14, 15, 2] |
| Publishing rate of the AI-agent factory | more than 175 models a week | C | XPRIZE announcement, from the founder's submission. | [4] |
| Team | 1 person | C | TrustMRR lists a team size of 1 person; the home page says "Hey, I am Lucas and I build Polyfork" and the footer "Built by Lucas Martinic". The pages often write "we". | [1, 2] |
| Outside funding | none found; a US$500,000 prize | B | TrustMRR lists the company as bootstrapped (founder-entered). XPRIZE announced on 26 September 2026 that Polyfork won the US$500,000 grand prize, which is prize money, not an investment for shares. It came after the August and September revenue. | [1, 4] |
| Visitors and Google impressions, last 30 days | 967 visitors; 2,267 search impressions | A | Google Search Console is connected to the profile; the visitor count is TrustMRR's snapshot. TrustMRR gives US$4.84 of revenue per visitor. | [1] |
| Time from launch to the October figure | 85 days, under 3 months | calc | Calculated: 1 August to 25 October 2026. The founded date is entered by the founder on TrustMRR. | [1] |

How to read this table: a grade A value comes from Stripe, the payment processor, through TrustMRR's connection to Polyfork's Stripe account. A grade B value comes from a third party that is not the founder. A grade C value is a statement by Polyfork or its founder, or a reading of the company's own pages and their archived copies. A "calc" value was computed on this page from the rows above it.

## When did Polyfork launch, and how did it grow? Timeline

From a store of 326 assets on 5 August 2026 to 2,634 assets and US$16,307 of verified revenue 85 days later. Each entry carries its grade.

- **1 August 2026** (grade C [1]): Founded date on the TrustMRR profile.
- **5 August 2026** (grade C [13]): The archived home page shows 326 assets, an MCP server with ten tools, Pro at US$10 a month and a Founders Club seat at US$100 once, with 85 of 100 seats left.
- **11 to 12 August 2026** (grade C [10, 14]): Blog posts on terrain programs and shader looks. The home page shows 457 assets, 115 new in the last 7 days, and 69 of 100 Founders Club seats left.
- **August 2026** (grade A [1]): US$8,917 of revenue, the best month so far.
- **27 August 2026** (grade C [8]): Characters v2: every humanoid becomes one engine called with a config, checked by automatic gates, the same lane that later builds US$9 custom skins.
- **September 2026** (grade A [1]): US$5,780 of revenue, 35.2% below August.
- **26 September 2026** (grade B [4, 5]): Polyfork wins first place and US$500,000 in the Build with Gemini XPRIZE. The announcement gives US$4,428 from 54 paying customers in the first month.
- **28 September 2026** (grade C [9]): A blog post compares two coding agents building the same game, one with Polyfork's catalogue and one without; the founder reports the farm game cost US$1.63 against US$3.08 in model fees.
- **3 October 2026** (grade C [15]): The archived home page shows 1,921 assets. The Founders Club has sold out; a Lifetime seat costs US$100 with 64 of 200 seats left, and Pro costs US$15 a month.
- **25 October 2026** (grade A [1, 2, 3]): 2,634 assets and 55 of 200 seats left. October revenue to date is US$1,610, and TrustMRR shows no active subscription.

## Why did Polyfork work?

### A model that is a program

A normal 3D model is a fixed mesh. A Polyfork model is a function, createAsset(), with declared settings, so the same file gives a short pine or a tall one, green or autumn coloured. The founder's argument on the home page is that an AI agent can only move a mesh around, but it can read a program's settings and change them. The knobs are also published as JSON, so agents need no human to explain them.

### AI agents write the catalogue

XPRIZE describes an AI-agent factory that publishes more than 175 models a week, and the company's posts describe a builder agent, a judge and automatic gates that test fit, seams and triangle budgets. The founder reads the kit requests before anything is built. The catalogue went from 326 to 2,634 assets in 81 days, which one person modelling by hand could not do.

### Built for coding agents

The site offers an MCP server, a REST API, llms.txt and prompt.txt, so Claude Code, Cursor or Codex can search the catalogue, fetch a matching part and place a whole kit with one call. Browsing needs no key. An agent can also start a purchase with the buy_plan tool, which the documentation says never charges anything itself.

### One-time seats paid the first months

The launch offer was a Founders Club seat: everything, forever, for US$100 once, limited to 100 seats. It sold out, and a Lifetime seat at the same price replaced it. The seat counter shows 145 seats taken, which would be at most about 89% of all revenue. This brought cash quickly, but each seat pays only once, and TrustMRR shows no recurring revenue.

## How it was built

These details come from Polyfork's own pages, its documentation for agents and the XPRIZE announcement. They are grade C unless the source is TrustMRR.

- Each asset is a self-contained ES module that returns a three.js group, with no loader and no waiting. The download is a zip with the module, a GLB, a README of every knob, the licence and an index.html.
- The same model is exported as GLB, FBX, USDZ and OBJ, with colours as named materials so that nothing arrives white in Unity or Blender. A Unity plugin to browse and import assets is on the founder's GitHub.
- Characters are one shared humanoid on a 22-bone skeleton with Mixamo bone names; no animation clips are bundled, and free Mixamo clips retarget onto it.
- The factory: a builder agent writes a config or module, automatic gates test it (fit-check, pose-check, seam-check and a 4,000-triangle budget for characters), and a judge model reviews it. The sources do not name the AI models used.
- An MCP server at polyfork.dev/mcp uses JSON-RPC 2.0 over HTTP POST, with no session setup. A Bearer key is needed only to download paid files.
- Stripe takes the payments, by TrustMRR's verification section and the privacy page. TrustMRR lists JavaScript for the frontend and PHP for the backend.
- Marketing on the sources: the founder's X account with 5,666 followers, ads on X (the privacy page describes their tracking), a Product Hunt page with 7 points, a weekly email of new assets, and the XPRIZE win.

Sources: [6, 7, 8, 16, 18, 19, 1, 17]

## Who tried this before Polyfork, and why did they fail?

Four companies in the Loot Drop database sold 3D content creation or 3D tools for the web before AI agents could write the content. The values below are from that database and are not independently verified.

The failed startups come from Loot Drop (https://www.loot-drop.io/), a free database of dead startups, retold here in our own words; Loot Drop describes its entries as AI-assisted summaries of public sources, so they carry no evidence grade.

### Modsy (2015 to 2022; funding US$72.7M; cause: Unit Economics)

- Promise: 3D room visualization that shows real furniture in a customer's own room before buying.
- Why it died: It was marketed as AI, but each room needed 2 to 4 hours of human designer work, so margins stayed low while it spent US$72.7 million.
- Record: https://www.loot-drop.io/startup/2521-modsy

### Sketchbox (2018 to 2021; funding US$1.5M; cause: Competition)

- Promise: Build VR and AR worlds without writing code.
- Why it died: Unity and Unreal offered stronger tools and larger communities, and VR headsets spread slowly.
- Record: https://www.loot-drop.io/startup/1028-sketchbox

### Famo.us (2011 to 2017; funding US$31M; cause: Competition)

- Promise: An HTML5 framework for smooth 3D-style animation in the browser.
- Why it died: Its own rendering engine had to be updated for every browser change, and simpler open-source frameworks such as React won the developers.
- Record: https://www.loot-drop.io/startup/133-famous

### MyMiniLife (2006 to 2010; funding US$120K; cause: Competition)

- Promise: Small virtual worlds that people placed on their blogs.
- Why it died: Simple web graphics lost to richer games, and Zynga bought it in 2009 and used the technology in its own products.
- Record: https://www.loot-drop.io/startup/361-myminilife

## What changed

Modsy is the closest lesson. It needed 3D models of real furniture and hours of human design work for every room, and the cost of that labour decided its fate. Polyfork has the same need for a large catalogue of 3D models, but AI agents write the models as code and automatic gates check them, so one founder added more than 2,000 assets in under three months.

The buyers also changed. Sketchbox and Famo.us sold tools to people who build 3D scenes by hand. Polyfork also sells to AI coding agents that build games and scenes from a sentence. Its catalogue is documented for them in llms.txt, prompt.txt and an MCP server, and its own test reports that an agent with the catalogue built a farm game for about half the model fees of an agent without it.

## Can one person copy Polyfork now?

**Verdict: The format, yes. The revenue model, not yet proven.**

Turning a product into code that agents can both write and read is a method others can use for icons, sounds, shaders or level designs. The revenue so far does not show that it lasts: one-time seats brought most of the money, and October revenue fell to about a fifth of August's daily rate. A US$500,000 prize gives the founder time to find out.

1. Choose content that can be expressed as a small program with settings, so an AI agent can produce many items and another agent can adjust them.
2. Build automatic checks for every item before it is published, and publish the failures too.
3. Document the catalogue for agents (llms.txt, an MCP server, a JSON schema of the settings) as well as for people.
4. Open with a limited one-time offer for cash, but plan the recurring plan that must replace it.

**What is still open.** The same agent-readable catalogue for other content types: 2D sprites and tiles, sound effects, UI components, or 3D parts in a higher-detail style.

## What could go wrong for Polyfork?

- **Revenue fell every month.** US$8,917 in August, US$5,780 in September and US$1,610 in October to the 25th. By daily average, October is 77.6% below August (calculated). In October about one US$100 seat sold every two days.
- **Almost no recurring revenue.** TrustMRR shows US$0 of monthly recurring revenue and no active subscription on 25 October 2026. The seat counter shows 145 lifetime seats, which would be up to about 89% of all revenue. Lifetime buyers get every future asset without paying again, and 55 seats are left before the offer closes.
- **Numbers that do not agree.** XPRIZE gives US$4,428 from 54 customers for the first month, while Stripe shows US$8,917 for August. The window of the XPRIZE month is not stated. On TrustMRR the last-30-days total (US$3,780), the sum of its daily rows (US$3,580) and a stored snapshot (US$4,780) also differ.
- **A side project with a prize.** The home page says the founder builds XR "for a living", and his GitHub profile lists him as an XR creative technologist. The US$500,000 prize is not revenue and is not in the Stripe figures.
- **Copies and competition.** The licence forbids reselling the files or using them to train a commercial asset generator, but the modules are readable source. Image-to-3D generators, such as Tencent's Hunyuan3D that Polyfork's own blog discusses, improve each year and could make a pre-built library less needed.

## Sources

1. TrustMRR: Polyfork, Stripe-verified revenue profile (data as of 25 October 2026) — https://trustmrr.com/startup/polyfork (grade A). Read the markdown profile (https://trustmrr.com/startup/polyfork.md) in full, once, on 25 October 2026. Grade A for the metrics in its Verification Sources section (Stripe API key, last synced 25 October 2026); the description, founded date, team size and funding are founder-written, grade C.
2. Polyfork home page, read 25 October 2026 — https://polyfork.dev/ (grade C). Read in full. Source for the product, the MCP tools, the asset count, the founder's note and the seat counter.
3. Polyfork pricing page, read 25 October 2026 — https://polyfork.dev/pricing (grade C). Read in full. Source for the plans, the Lifetime seats, the end of the Founders Club and the reduced price.
4. XPRIZE, 26 September 2026: Polyfork wins the Build with Gemini XPRIZE for building a scalable AI business in 90 days — https://www.xprize.org/news/polyfork-wins-the-build-with-gemini-xprize (grade B). Read in full. The organizer's announcement. Source for the prize, the founder's country, the AI-agent factory and the first-month figures (which are the founder's submission, grade C).
5. Pulse 2.0: Polyfork Wins $500,000 Build With Gemini XPRIZE After Building AI-Powered 3D Model Business In 90 Days — https://pulse2.com/polyfork-wins-500000-build-with-gemini-xprize-after-building-ai-powered-3d-model-business-in-90-days/ (grade C). Read in full. It retells the XPRIZE announcement and adds no new figure.
6. Polyfork licensing page, last updated August 2026 — https://polyfork.dev/licensing (grade C). Read in full. Source for the licence, the rule against training a commercial generator, and the factory orders.
7. Polyfork llms.txt, the catalogue written for AI agents — https://polyfork.dev/llms.txt (grade C). Read in full. Source for the skeleton, the skin and steering prices, and the formats.
8. Polyfork blog, 27 August 2026: Characters v2, our 3D characters are programs now, not meshes — https://polyfork.dev/blog/characters-v2 (grade C). Read in full. Source for the character engine, the gates and the builder-and-judge lane.
9. Polyfork blog, 28 September 2026: Build a whole low-poly 3D world in three.js with one call — https://polyfork.dev/blog/a-kit-is-a-place (grade C). Read in full. Source for the kit loader and the founder's test of two coding agents.
10. Polyfork blog, 11 August 2026: Hunyuan3D WorldClaw and the 3D asset supply problem — https://polyfork.dev/blog/worldclaw-and-the-3d-asset-supply-problem (grade C). Read in full. The founder's view of generated meshes against a parametric library.
11. Polyfork custom skin page — https://polyfork.dev/skin (grade C). Read in full. Source for the US$9 skin and what the company refuses to build, including real people without their consent.
12. Polyfork: Steer the factory — https://polyfork.dev/steer (grade C). Read in full. Source for the kit ideas board and the founder's review of each idea.
13. Internet Archive copy of polyfork.dev, 5 August 2026 — https://web.archive.org/web/20260805111834/https://polyfork.dev/ (grade C). Read the archived home page in full. Source for the early asset count, prices and seat counter.
14. Internet Archive copy of polyfork.dev, 12 August 2026 — https://web.archive.org/web/20260812162920/https://polyfork.dev/ (grade C). Read the archived home page in full.
15. Internet Archive copy of polyfork.dev, 3 October 2026 — https://web.archive.org/web/20261003062730/https://polyfork.dev/ (grade C). Read the archived home page in full. Shows the Lifetime plan after the Founders Club sold out.
16. Polyfork: connect the catalogue to your AI agent — https://polyfork.dev/agents (grade C). Read the sections on the MCP tools and setup.
17. Product Hunt: Polyfork — https://www.producthunt.com/products/polyfork (grade C). Read the product page and the founder's launch comment. It shows 7 points and 37 followers.
18. GitHub: lucas-martinic — https://github.com/lucas-martinic (grade C). Read the profile and the repository list. Source for the founder's job title, the Unity plugin and his other projects.
19. Polyfork privacy page, last updated August 2026 — https://polyfork.dev/privacy (grade C). Read in full. Source for Stripe payments and the ads on X.
20. Loot Drop database export, failed startups — https://www.loot-drop.io/ (grade C). Local export. AI-written summaries, not independently verified. Source for the failed predecessors.