What is vibe coding?

The definition

In a February 2025post on X, Andrej Karpathy described a workflow he called “vibe coding”: rather than reading or authoring the code himself, he gave the AI “vibes” — corrections, ideas, refactors — and accepted whatever the model produced as long as the app worked. He went so far as to say he had “forgotten the code even exists.” The framing spread inside weeks.

The underlying practice was not new — agentic coding had been growing since 2023 — but the name was. Vibe coding crystallised a specific stance: the human treats the code as an artefact owned by the model, intervening only when the running application falls short of intent.

As of 2026, “vibe coding” covers a range of workflows that share three traits: (1) the human writes natural-language goals rather than source code, (2) the agent has tools to read, write, and execute, and (3) the human reviews running behaviour more than diffs.

How it differs from related practices

The term is often confused with three neighbours. Each shares some mechanics but differs on what the human pays attention to.

A short timeline

1. 2021GitHub Copilot launches as a Visual Studio Code extension, with OpenAI Codex behind it. Sets the template for in-IDE AI autocomplete.
2. Mar 2023GPT-4 release. Reasoning quality crosses a threshold that makes long-form coding from English plausible for the first time.
3. 2023Auto-GPT, BabyAGI, smol developer popularise autonomous agent loops. Most are toys; the pattern sticks.
4. Late 2023Cursor and Aider establish the two dominant modes — IDE-native and CLI-first agentic coding.
5. Mar 2024Cognition launches Devin, the first fully managed cloud-sandbox software engineer. Long-horizon tasks become possible.
6. Nov 2024Anthropic introduces Model Context Protocol (MCP). Coding agents gain a standard way to expose tools and read context.
7. Late 2024Browser-based prompt-to-app builders take off — Lovable, Bolt, v0, Replit Agent. Non-engineers ship products without writing code.
8. Feb 2025Karpathy posts the “vibe coding” tweet. The term becomes the umbrella label.
9. 2025–2026Claude Code, Windsurf, OpenHands, Aider, Roo Code, Kilo Code, and dozens more enter the market. MCP becomes a de-facto standard. First entire companies are founded with no full-time engineers, shipping their products via vibe coding.

Representative tools

The ten most-cited vibe-coding tools as of 2026, ordered roughly by how often they appear in answers to “what should I use to vibe-code an app?”

1. 01
   Cursor

   AI code editor forked from VS Code with built-in agent, multi-file edits, and tab-completion.

   →
2. 02
   Claude Code

   Anthropic's official CLI agent for Claude — runs in the terminal, edits files, executes commands, and ships PRs.

   →
3. 03
   Lovable

   Browser-based AI app builder that generates full-stack React + Supabase apps from natural language.

   →
4. 04
   Bolt.new

   StackBlitz's prompt-to-app builder running on WebContainers — full Node.js stack runs in your browser tab.

   →
5. 05
   v0

   Vercel's AI app builder optimised for Next.js, shadcn/ui, and Vercel deployment.

   →
6. 06
   Aider

   Open-source CLI coding agent that pair-programs in your terminal and commits to git automatically.

   →
7. 07
   Windsurf

   AI code editor by Codeium with Cascade agent for autonomous multi-step coding tasks.

   →
8. 08
   Replit Agent

   Replit's autonomous build agent that creates, debugs, and deploys apps inside the Replit workspace.

   →
9. 09
   Cline

   Open-source autonomous coding agent for VS Code, formerly known as Claude Dev.

   →
10. 10
    Devin

    Cognition's autonomous software engineer — runs tasks in a managed cloud sandbox with browser, terminal, and editor.

    →

Browse the full directory at /tools.

What you need to start

The minimum kit for a serious vibe-coding session in 2026 is small: one agent, one model subscription, one project sandbox, one feedback loop. Specifically:

• An agent. Cursor or Windsurf if you want an IDE; Claude Code or Aider if you prefer a terminal; Lovable, Bolt, or v0 if you want to skip local setup.
• A frontier model. Claude Sonnet 4.6, GPT-5, or Gemini 2.5 Pro at minimum. Smaller models still write code, but their reasoning over multi-file edits is materially worse.
• Fast tests. Vibe coding works only when you can tell within seconds whether the agent broke something. Tight type-checking, hot-reload, and unit tests are the lifeline.
• Version control with rollback. Commit small, commit often. The agent will sometimes go off the rails; git is the safety net.
• A spec, even a small one. The single biggest quality determinant is how well you describe the goal up front. Two paragraphs of plain English beats a one-line prompt by far.

Limits and criticism

The term has detractors. Three lines of criticism keep coming back:

1. You can't debug what you didn't write. When the system breaks in production at 3 a.m., the human who never read the code is poorly placed to fix it. Senior engineers tend to read more code than the rhetoric suggests.
2. Quality drifts as the codebase grows. Agents handle the first 1,000 lines brilliantly and degrade past 10,000 lines. Architecture decisions made invisibly compound.
3. Security and compliance are not vibes. Code with regulatory implications — payments, auth, healthcare — still needs human review at every change, regardless of how polished the output looks.

The honest reading is that vibe coding is a great mode for some tasks and a poor mode for others. Treat it as a tool, not an identity.