How Hexlet automates business processes beyond development
Как Хекслет автоматизирует бизнес-процессы и ускоряет не только разработку
B · Useful, limited Hexlet founder Kirill Mokevnin's first-hand account of wiring a small company's SaaS stack to Claude and Codex through MCP, with concrete mechanisms across HR, content and marketing; numbers are his own and the '10x' is an estimate.
Summary based on the full transcript·Original: about 62 minutes to listen
In short
A solo episode (18 October 2026) of Организованное программирование (Organized Programming) by Kirill Mokevnin, founder of Hexlet, a Russian online programming school with about 40 staff plus contractors. Over roughly two months the company connected every SaaS tool it uses to Claude, Codex or ChatGPT through MCP. He describes what changed in hiring, content support and marketing, which SaaS products he expects to lose, and why in-product AI assistants lose to agents with cross-system context.
Key numbers
- 300 to 30-40candidates in the hiring funnel screened to a shortlist in one day through an agent, without opening the HRM
- 40,000lines in the PR that added Hexlet's own MCP server, mostly 300+ near-identical admin tools
- 5 to 1people needed for the practice-content support area after agent-driven ticketing and fixes
- ~95%share of the micro team's hundreds of monthly tickets that are back-office rather than user-facing
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- SaaS that survives is tied to regulation, payments, deep integrations, contracts or risk; single-user tools lose. Tilda shipped an MCP server and Hexlet, which runs hundreds of landing pages, adopted it at once.
- In-product assistants lack cross-system context: categorising a lump-sum payment payout in a finance tool needs an agent that can see the payment provider, banks and the finance system together.
- Hiring: he screened candidates from the Talantix HRM through his own Codex session driving the browser, because Talantix has no MCP. In one day: 300 people in the funnel, 30 or 40 shortlisted, the rest rejected, without opening the HRM himself. None of the popular Russian HRM systems he checked has MCP, and most charge for API access.
- Hexlet's own MCP: a 40,000-line PR, mostly more than 300 near-identical tools over admin operations, merged quickly with a few fix iterations after.
- Course feedback is turned into tickets through MCP, and a Codex skill has run for three weeks filing and fixing small content issues. An area once staffed by five people is now handled by one.
- Non-technical staff built dashboards that re-pulled all data on every refresh and exhausted $20 plans; the fix he teaches is to ask the model how a task should be done before doing it.
- Jobs: people doing routine operations without analysis or decisions are most exposed. His "ten times more per person" is stated as an estimate for non-engineering functions, not for software delivery. He expects much less hiring and consolidation among agencies and outsourcers.
How they got it
The founder's own experience at his company, narrated with examples. No measurement method; figures are his recollection.
Limits
- Self-reported, one small company, with two sponsor reads (his consultancy and Hexlet courses).
- The 10x figure is an explicit estimate.
- Agent-based screening of candidates raises fairness and data-protection questions that the episode does not address.
Why it matters for companies
The account covers the whole company, not only developers, and names the mechanism: MCP connectors as the integration layer, with each person's own permissions. The HR example is pointed: if screening runs through an agent outside the HR system, the HRM loses its role, and HR-tech vendors without MCP become the obstacle. Firms should decide who audits agent-made hiring decisions before staff start doing this on their own.
Key passages in translation
Russian original
воронке 300 человек, из них я отобрал 30, по-моему, или 40. Но остальные ушли в отказ. И я при этом ни разу не зашёл на сам
Three hundred people in the funnel; of those I picked 30, I think, or 40. The rest were rejected. And I never once went into [the system] itself.
Russian original
А что это за 40.000 строккода? А это просто МCпишки. То есть там больше 300 инструментов.
So what are those 40,000 lines of code? They are just MCP tools. There are more than 300 tools there.
Russian original
классическая производительность труда, да, один человек может делать там в десяток раз больше.
Classic labour productivity: one person can do about ten times more.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com