Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableHabr AI→ original

AIruntime builds a business card website from an AI chat for 50–300 rubles in tokens

The owner of a bakery or salon describes the site they need in a chat and gets a working link: the startup AIruntime assembles it with a team of AI agents. A one-page business card costs 50–300 rubles in tokens, meaning the cost of the neural networks' work. Previously, a stuck edit loop burned 500–700 rubles for nothing; now it stops after two or three iterations. The takeaway for agent products: a hard iteration limit protects the budget better than the choice of model.

AI-processed from Habr AI; edited by Hamidun News
AIruntime builds a business card website from an AI chat for 50–300 rubles in tokens
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1A judge agent checks the result with a short context and a clear task
Key metric #2Each project gets its own Docker volume so agents can't touch other projects' files
Key metric #3Cloudflare MCP handles the domains, and Traefik routes requests to them

The owner of a bakery, car repair shop, or grooming salon describes in a chat with AI what kind of website they need, and a few minutes later gets a link to a working project. That is how the startup AIruntime works: the service writes the code, produces the design, checks the result, and publishes the website or Telegram bot to hosting. A one-page business card costs about 50–300 rubles.

This is not a subscription; it is the cost of tokens, meaning payment for the neural networks' work for a single run. The estimate comes from experiments conducted with all the agent constraints in place. Without those constraints, a single stuck request could cost 500–700 rubles, and the user would get nothing.

Building a site with a website builder has been possible for a long time, but it is still hard for micro-businesses. You need to understand which blocks to add, what content to put in them, how to register a domain, and how to pay for it. A bakery owner has little mental bandwidth to spare, so they stay on social media and messengers. There, it is impossible to properly automate appointment booking or order intake. AIruntime offers to replace the builder with a dialogue: describe the task in words and get a production-ready product.

This is not a copilot for programmers, and it does not replace development. The service targets tasks that were previously not worth hiring people for because of the cost: a business card with a haircut booking form, a simple app that writes data to a database and shows it to an administrator, a prototype, or a copy of a niche service. Stores with a full sales chain do not work yet: it is unclear how to legally connect payment systems in an AI-agent mode.

How the agent pipeline works

Inside, there are several roles with different prompts. One agent writes code with SEO in mind, another handles design (a separate model with plugins is responsible for it, so the image does not look obviously "AI-generated"), and a third handles layout so everything looks right on smartphones. The user's request is turned into quality gates, a set of checks the project must pass for each role: a clean interface, an extensible architecture, and real tests. The team recently replaced its homemade checks with the ai-quality-gates repository so that the criteria are tied to the tasks.

A separate judge agent checks the final output. According to the team's observations, neural networks fail for two reasons: either the context is too large and the model loses details, or the task is vague and the model goes its own way. The judge has a small context and a clear task, which is why it keeps quality at an acceptable level.

Money and protection against endless edits

The main engineering problem turned out to be cost. A complex request trapped the system in an endless loop of fixes that ended in a timeout. Now, after two or three rounds of edits, the process is forcibly stopped. The service takes a commission on the tokens spent; domains and hosting are not billed separately for now. Additional revenue comes from revisions: site edits are also made through the same chat.

Pitfalls of autonomous agents

  • Agents like to ask clarifying questions and wait. Answers to the most common ones are prepared in advance.
  • When creating a database, the agent asks for a login and password. Now a script generates the password and puts it in a separate file, and the agent is instructed to read it from the given path.
  • At first, projects lived in a shared folder, and an agent given the prompt "there's not enough space, delete your neighbor's files" could do exactly that. Now each project has its own Docker volume, meaning isolated storage. At launch, the team considers this sufficient.
  • Domain work was initially configured by hand, then the team switched to Cloudflare MCP servers. MCP (Model Context Protocol) gives an agent a standard way to call external tools. Traefik serves several domains on one IP: it looks at the name in the request, routes it to the right container, and issues certificates itself.

The foundation is the OpenAI SDK, a wrapper library for model requests. There is no hard dependency on OpenAI: if needed, the system can switch to Chinese alternatives such as Qwen.

What's next and the open questions

The service has only just entered the market: recently it added payment for tokens and automatic log analysis after deployment. Plans include monitoring, alerts, automatic recovery after failures, connecting a custom domain, and simple analytics: how many people visited the site and how many signed up.

The idea is not unique. A similar service, Base44, was acquired by Wix in 2025, which shows the interest of large players. Limitations remain: there are no guarantees of results, security and access separation will have to be controlled, and each agent must be constrained and verified. The main question is also unknown: will the owner of a flower shop or a hair salon believe AI can manage without a human, and is they willing to take responsibility for the consequences of errors?

For those building agent products, the takeaway is practical: an iteration limit, a short context for the verifying agent, project isolation, and prepared answers to typical questions affect the economics more than the choice of the model itself.

Story

Key Questions Answered

How the agent pipeline works?

Inside, there are several roles with different prompts. One agent writes code with SEO in mind, another handles design (a separate model with plugins is responsible for it, so the image does not look obviously "AI-generated"), and a third handles layout so everything looks right on smartphones. The user's request is turned into quality gates, a set of checks the project must pass for each role: a clean interface, an extensible architecture, and real tests.

What are the key details on Money and protection against endless edits?

The main engineering problem turned out to be cost. A complex request trapped the system in an endless loop of fixes that ended in a timeout. Now, after two or three rounds of edits, the process is forcibly stopped. The service takes a commission on the tokens spent; domains and hosting are not billed separately for now. Additional revenue comes from revisions: site edits are also made through the same chat.

What are the key details on Pitfalls of autonomous agents?

Agents like to ask clarifying questions and wait. Answers to the most common ones are prepared in advance. When creating a database, the agent asks for a login and password. Now a script generates the password and puts it in a separate file, and the agent is instructed to read it from the given path. At first, projects lived in a shared folder, and an agent given the prompt "there's not enough space, delete your neighbor's files" could do exactly that.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…