What is vibe coding and how to start in one evening

Vibe coding is a way of making software where you describe the task in ordinary words, an AI model writes the code, and you check the working result instead of reading the lines. The term comes from Andrej Karpathy, a co-founder of OpenAI. Below: the definition and the origin of the term, the tools, a first project for one evening, and the limits of the method that courses tend to skip.

Vibe coding in plain words

In ordinary development you think, write code, run it and fix it. In vibe coding you say "I need a habit tracker, keep the data in the browser". The agent creates the files and runs them, you look at what came out and correct it in words: "the delete button doesn't work".

The name is spelled several ways: vibe coding, vibecoding, vibe-coding. The meaning is the same. Karpathy described the extreme case, where the code is not read at all. In everyday talk the same phrase also covers more careful work with an AI agent, so it helps to tell three modes apart:

The line runs along the price of a mistake. A broken habit tracker upsets only you, so you can skip reading the code. If the program takes money or stores other people's data, the code always gets read.

Where the term comes from

Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, wrote on X on 2 February 2025 about a new kind of programming in which you "fully give in to the vibes" and forget that the code even exists (Wikipedia article on the term). Hence the name: you work from the feel of the result, not from knowledge of the details.

The term spread fast. In March 2025 Merriam-Webster put it in its "slang & trending" section, and on 6 November 2025 Collins named vibe coding its word of the year.

Tools and models for vibe coding

Every tool runs the same loop: you describe the task, the agent reads the project folder, creates and edits files and runs commands, and you look at the result and refine. What differs is the shell and the price.

Cline, Aider and Continue are covered in Free vibe coding: which tools are real. Installing OpenCode and connecting models is described in a separate guide.

The result depends on both the shell and the model under it. We would not pick a model from a slide deck: take two or three and compare them on your own task. The ranking by scenario and the testing method are in The best AI for programming, and a selection aimed specifically at vibe coding is in The best AI for vibe coding.

Your first project in one evening, step by step

Take a habit tracker: one HTML page, no servers or sign-ups, data in the browser's localStorage. Any mistake here is harmless, and the result is visible at once.

Step 1. Install the tool. For Claude Code, the commands come from the official documentation:

curl -fsSL https://claude.ai/install.sh | bash

irm https://claude.ai/install.ps1 | iex

The first line is for macOS, Linux and WSL, the second for Windows PowerShell. Check with claude --version. You need macOS 13 or newer, Windows 10 1809 or newer, or Ubuntu 20.04 and newer, with at least 4 GB of RAM.

Step 2. Create a folder and a repository. Before the first claude command you need access to a model: an Anthropic subscription or an API key. Setting it up through codexkey is covered in the next section.

mkdir habit-tracker && cd habit-tracker
git init
claude

Step 3. Describe the task, with a plan. Ask for a plan in the first request, not for code:

Build a habit tracker on a single HTML page with no libraries and no server.
Habits can be added and deleted, each day gets a check mark,
data is kept in localStorage. Show the streak of consecutive days.
First lay out a five-point plan, do not write any code yet.

If the plan works, say "go". If not, adjust it in words: that is cheaper than fixing finished code.

Step 4. Work in small steps. After every change, open index.html in the browser. Check four things: a habit can be added, a check mark can be set, deletion works, and after reloading the page the data is still there. One request changes one thing.

Step 5. Save what works. It works: git add . && git commit -m "Working tracker". The next edit broke everything: git restore . brings back the last live version.

Step 6. Add tests and ask what is left unchecked.

Move the streak calculation into a separate function and write tests for it:
a skipped day, a month change, a leap year. Then explain what each
function does and list what you did not check.

That last sentence is always needed. Asked directly, the model usually names its weak spots, but do not trust the list to be complete: treat it as a pointer to where you should look yourself.

By our estimate the evening breaks down like this: 15 minutes for installation, 30 for the plan and a first version, about an hour of iterations, 30 minutes for tests and 15 for reading what you got. The timing depends on the task.

Vibe coding through codexkey: key, address and the price of an evening

codexkey gives API access to models; it does not sell Claude or ChatGPT subscriptions. Claude Code needs a sk_cdx_... key and an address without /v1. The settings block from the Claude Code guide, in the file ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.codexkey.online",
    "ANTHROPIC_AUTH_TOKEN": "sk_cdx_YOUR_KEY",
    "ANTHROPIC_MODEL": "claude-sonnet-5[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5-5[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-sonnet-5",
    "CLAUDE_CODE_SUBAGENT_MODEL": "claude-sonnet-5"
  }
}

You do not need to sign in to an Anthropic account. The balance can be topped up with a Russian bank card, through SBP (the Russian Faster Payments System) or with crypto. Cursor connects on the same principle, but according to the codexkey guide, Cursor does not accept custom endpoints during its trial period.

How much does an evening cost? Suppose the agent reads 2M input tokens and writes 150 thousand output tokens over three hours (an assumption, yours will differ; caching is not counted). The prices come from the codexkey price list as of 29 October 2026; for current ones see the models page:

The formula: input in millions of tokens times the input price, plus output in millions times the output price. For Sonnet 5.5 that gives 2 × 0.52 + 0.15 × 2.6 = 1.43 dollars. Sonnet 5 from the settings block above is priced the same as Sonnet 5.5.

A free start, and what about courses

You can vibe code for free, but "free" here means "with limits". What exists as of 29 October 2026:

• Cursor: a free plan with a limited number of agent requests.
• OpenCode console: the list of free models includes DeepSeek V4.1 Flash Free and GLM-5.3-Flash Free. Privacy terms differ between models, so read about each one on the documentation page.
• The codexkey catalog: the models deepseek-v4.1-flash-1017-free, minimax-m3-free and qwen3.8-flash-free cost $0 per token. Limits and terms are on the models page.
• Claude Code: there is no free plan, so you need a subscription or an API key.

Do you need courses? Often not. What to learn instead (git, reading code, tests, key safety) is covered in Vibe coding courses: are they worth it. For 1C, vibe coding works differently, with details in the 1C article.

Risks and limits of vibe coding

Two studies are worth knowing before your first working project.

Speed. In a 2025 METR experiment, 16 experienced developers worked through 246 tasks in their own large repositories. With AI the tasks took 19% longer, although the participants themselves believed they had sped up by 20%. According to METR's update of 24 February 2026, the speedup for late 2025 was estimated at 18% for returning participants and 4% for new ones, but in both cases the confidence intervals include zero. The lab itself called the result an unreliable signal because of how participants were selected, and added that by early 2026 developers are most likely speeding up more (METR post). For a beginner the takeaway is modest: the feeling of speed and real speed diverge, so measure on your own tasks.

Security. A Veracode report from October 2025: code generated by more than a hundred models in Java, JavaScript, Python and C# contained risky vulnerabilities in 45% of tests (Veracode). That does not mean all generated code is full of holes. It means "works" and "is safe" are checked separately.

The rules we advise you to keep:

1. Keys and passwords stay out of the repository and out of prompts.
2. Payments, authorization and other people's personal data are not entrusted to code nobody has read.
3. Do not start without git and tests. A way to roll back matters more than a clever prompt.
4. Give the agent no rights you are not ready to lose: in July 2025 the founder of SaaStr reported that a Replit agent deleted a production database despite an instruction to change nothing.

FAQ

What is vibe coding in plain words?

A way of making software by describing the task in ordinary words: an AI model writes the code, and you run the result and correct it with new words. The term appeared on 2 February 2025 with Andrej Karpathy, who meant a variant where the code is not read at all.

Why is it called vibe coding?

From the English word vibes: you work from the feel of the result, not from knowledge of the details. Karpathy wrote that in this mode you give in to the vibes and forget the code exists.

How do I start vibe coding from zero if I can't program?

A simple one-evening project is within reach if you go in small steps, keep git and check the result by running it. For anything that stores other people's data or takes money, you need to be able to read code or bring in someone who can.

How can you make money from vibe coding?

There are no guarantees, and we do not name sums. Realistic options: automations and internal tools for a specific client's task, prototypes for testing an idea, small websites. You, not the AI model, answer for the result and its security.

Which AI model is best for vibe coding?

There is no universal winner. Start with Claude Sonnet or a mid-tier GPT, compare against one cheaper and one stronger model on your own task, and count the cost of a solved task rather than the price per token. The testing method is described in The best AI for programming.

Can you vibe code for free?

Yes, while the limits last: Cursor has a free plan, the OpenCode console has free models, and the codexkey catalog has three models at zero cost per token. Claude Code has no free plan. Free tiers are limited in requests, and for some models your data may be used to improve the product.