One Sunday, 202 sub-agents ran: a Russian product-management software that had been around for 8 years put Claude Code inside

catalog.app research report

Data card: /home/ljr/fanout/why/cards/web_catalog.app.json (web channel, C2 pool). Research date 2026-11-07.

One line

catalog.app is a PIM hub built by a Belarusian team for large sellers and distributors in Russia, Belarus and Kazakhstan (supplier price-list ingestion → fuzzy matching against its own 80M-product library → filling product cards → pricing → pushing to 7 platforms including Ozon, WB, Yandex.Market and Kaspi, plus 1C). It has been running for 8 years. In summer 2026 it wrapped Claude Code / OpenCode into a built-in AI agent that does the work directly, and the ranking climb roughly coincides with that launch. But the climb mostly reflects existing clients using it much harder and more machine traffic, not a sudden influx of new users.

1. What it is

• Official site catalog.app: Russian-language, the product is a web SaaS. The login page /Home/Login and the landing page share one origin, so the CrUX and Tranco traffic mostly comes from the product itself, not marketing pages. There's no app.
• Legal entity: ООО «Каталог.АПП» (TIN 9709087833, OGRN 1227700771650, registered in Moscow in 2022), disclosed under Russian Ministry of Digital Development order no. 511 (catalog.app/mintsifry-511). The site also lists a Belarusian phone number +375, server IP 93.125.82.102 and by./minsk. subdomains, and the team's roots are in Minsk. Founder / lead developer Yury Loskat (Habr username Razoomnick, vc.ru id 521033).
• Scale claims (from the site): running in production for 8 years; product library of 80M cards (the new Habr article says 100M), 300M images, 7M barcodes; 8M price lists processed to date; target clients are sellers, distributors, retail chains and brands with 10k to 3M+ SKUs. No public pricing; you need a demo first (there's a demo site at ru.demo.catalog.app).
• 2026-05 To get around Russia's network blocks, it split out three entry points: ru.catalog.app (Russia), by.catalog.app (Belarus) and de.catalog.app (Germany). The main catalog.app site switches between them "as needed" (source: official Telegram channel t.me/httpscatalogapp, 2026-05-05 post). Certificate transparency logs also show subdomains such as browser-proxy, library-search and manual-parsers (crt.sh).

2. The data: what actually went up

How to read these numbers

• Real browser traffic (CrUX) rose only one tier (top 1M → top 500k), placing it alongside mass-market seller tools like SelSup and XWAY.
• Tranco and Radar, however, put it ahead of mass-market seller tools: more than double SelSup, level with pimcore.com, and Radar reached the top 100k. A B2B PIM that sells only to large clients on demo-based contracts likely has a customer count in the hundreds (unverified). A DNS ranking this far above the browser ranking suggests a large share of the rank climb is machine traffic / usage intensity (API, MCP calls, browser-extension callbacks, image hotlinking), not more visitors.
• Timeline: Tranco began climbing from the trough in early September, with the steepest slope in mid-October. Since Tranco is a 30-day rolling average lagging 1–2 weeks, the real inflection falls around mid-to-late August, which matches the CrUX tier jump in August.

3. Why it took off: the AI agent launch was the inflection

The founder published several long posts on vc.ru and Habr, and the timeline is clear:

• Background: Three earlier attempts to add AI to the product (generating descriptions, filling attributes, categorization) were only "small features". The GPT description generator built in 2023 was used by only 20% of clients (vc.ru 2026-09-23 post).
• What they did: The founder began heavy use of Claude Code in 2025 December. One evening he built a prototype, giving Claude Code the platform's data-query tools, and in 1 hours assigned customs codes (ТН ВЭД) to 100k products, with differences from manual results of only a few tenths of a percent. Full development followed: every page embeds a chat box, the backend is a standalone agent runner executing Claude Code / OpenCode, calling hundreds of tools through the platform's own MCP server, and able to "look at pages like a user." To polish the agent until it passed clients' probing questions, the team deliberately delayed the release by a month or two (Habr 2026-10-25 "Как теперь устроена наша PIM-система…" habr.com/ru/articles/1098978/).
• Results (per the founder): "Now it's several thousand conversations a week"; after launch he met with clients specifically about the agent 50 times or more; support requests dropped to about a third of previous levels. An Ozon seller with a 130k-SKU catalog opened 32 chats on one Sunday; the agent dispatched 202 sub-agents in parallel and assigned 10,368 customs codes in 41 minutes. Another client dropped in 292 rows of a new supplier price list and had 291 products built and priced within half an hour (vc.ru 2026-09-23 vc.ru/life/3153804, 2026-09-25 vc.ru/life/3158796).
• Why this shows up in traffic:
  1. The chat script is attached to every page and, in the founder's words, "uses plain http requests, switching to websocket later," which amounts to polling: requests keep flowing as long as a client's page is open;
  2. Agent tasks run for hours and spawn sub-agents, and every step calls back to catalog.app via MCP;
  3. The 2026-07-02 browser extension was upgraded to "use AI to auto-generate parsers for any website, and auto-traverse the entire site," sending a callback to catalog.app on every page the user's browser loads (Telegram 2026-07-02 post);
  4. Since 2026-05, images on the ru entry are served from its own servers instead of a CDN.

All four factors fall in the same period and each would greatly increase DNS / request volume without a proportional increase in CrUX page visits, which fits the observation that Tranco and Radar rose far more sharply than CrUX.

• Where users come from and why they stay: In the founder's 2024-12 Habr article "Мифические 250 человеко-месяцев," he says acquisition runs on client referrals and articles where the team shares experience; he often hears clients say they found the product by "reading the Habr articles" (the one on bypassing CAPTCHAs drew 41k reads, the one on fuzzy search 11k). Clients stay because switching costs are high: supplier price-list onboarding, marketplace integrations, and pricing DSL setup are all things clients have already spent time configuring. The self-built 80M-product catalog and 8 years of accumulated handling of platform API quirks (the founder's example: docs list 20 fee types, but they actually encountered 50) are hard for others to replicate. The AI agent is a "doing layer" added on top of this foundation, and users go deeper the more they use it.

4. Growth tactics (evidence)

5. Competitors

Bottom line: very few players in Russian-speaking markets run the full chain "supplier price list → matching → product card → pricing → multi-platform" and also own a large product catalog. Embedding a general-purpose Claude Code-style agent into a PIM, with MCP letting it edit the customer's Ozon back office directly: I found no second one.

6. Differentiation

1. Data and matching algorithm: 80M to 100M product cards plus 300M images; fuzzy matching has been measured continuously since 2018 (auto-match 84%, error rate 0.1%). Built over 8 years from 8M price lists.
2. The whole chain in one system: collection, matching, enrichment, pricing (Scriban DSL), publishing, management accounting. Competitors usually cover only one or two of these steps.
3. AI agent on top of a complete platform: the agent can call every function and dataset (customs codes, product catalog, platform APIs), so it can actually finish the job. A plain ChatGPT or card generator can't. This is what the founder means by the decisive trait of the one attempt out of four that became a product.

Is this why it took off? Points 1 and 2 are why it has lasted 8 years and why customers stick around, but they didn't make it "take off". Point 3 and the ranking inflection in summer 2026 line up in time, and are the most likely cause of this recall signal. But what it mainly amplified is usage intensity and machine requests from existing customers; there's no evidence it brought in many new customers.

7. Visibility and copycats

• Revenue is not public; no TrustMRR listing, no press coverage; no mentions on X; the Telegram update channel has only 60 subscribers; the new Habr post has 7.7K reads. This sits in a small B2B circle in the Russian-speaking world, and is largely unknown in the English and Chinese-speaking worlds.
• Copycats: I haven't seen anyone copy "a general-purpose agent built into the PIM" yet. Brandquad and others are building AI card features, but they're still single-point functions.

8. What to borrow

• Embed Claude Code / OpenCode as the executor inside an existing B2B SaaS: build your own MCP tool layer (hundreds of tools), a layered documentation wiki, a "look at the page" ViewEngine, and an agent runner. Don't build the agent yourself; the moat is in the tools and data.
• Polish until it passes the customer's "probing questions" before shipping: better to be one or two months late than to lose trust on first impression.
• Hand high-value deterministic tasks to the agent (customs codes, required fields, deduplication, creating products from price lists), backed by the principle of "don't fill it in if unsure, leave it for a human to review" plus one-click rollback.
• The same logic of "price list → catalog matching → multi-platform listing and pricing" can be transplanted to distributor scenarios on Amazon / TikTok Shop / Mercado Libre / Shopee. The hard part is building the product catalog yourself.

9. Still unclear

• Customer count, revenue, and prices are not public; whether the AI agent brought in new contracts is only supported by a self-reported "50 multiple customer meetings".
• How much of the ranking climb comes from machine traffic (API / MCP / extension callbacks / images) versus human traffic can't be separated. Tranco and Radar are far higher than CrUX; this is an inference, not a measurement.
• The exact launch date of the AI agent: inferred from "postponed by one or two months", "read chat for a month", and the late-September post, roughly 7–August; not stated in the official changelog.
• The ECOM Expo'26 exhibition and the moySklad / cnews review rankings mentioned in the recall card have not been independently verified.