The rise of vibe coding

In 2024, large language models (LLMs) took a quantum leap in code generation, turning what was once a futuristic dream into a daily reality for developers. From novices hacking together their first apps to seasoned engineers architecting complex systems, these advancements leveled the playing field. Yet, it wasn’t all smooth sailing—while LLMs churned out impressive code snippets, they occasionally spat out non-functional API calls or quirky syntax, leaving developers to sift through the brilliance and the blunders. Still, the stage was set for something extraordinary, and by 2025, AI coding tools had evolved into indispensable allies, ushering in a new era of “vibe coding”—a term coined by AI luminary Andrej Karpathy that captures a seamless, intuitive, almost improvisational dance between human creativity and machine precision.

Karpathy, a computer science PhD from Stanford, co-founder of OpenAI, and former head of AI at Tesla, dropped the phrase “vibe coding” in a now-iconic X post on February 2, 2025. He wrote, “There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs (e.g., Cursor Composer w Sonnet) are getting too good.” With a career that’s seen him pioneer neural networks for self-driving cars and shape ChatGPT’s foundations, Karpathy’s not just some armchair theorist—he’s a heavyweight whose playful yet profound idea took off like wildfire, landing in the Merriam-Webster Dictionary as a “slang & trending” term by March 2025. His vision? Coding as a conversation, where you toss ideas at an AI—sometimes via voice—and it spins up functional code while you sip coffee.

TL;DR: In 2025, AI coding tools like Cursor, Replit, Windsurf and V0, powered by LLMs like Claude 3.7 and Grok 3, are turning pros and newbies into “vibe coders.” Creativity trumps syntax as 40M+ users hack apps without coding know-how, while LLMs fight IDEs for supremacy.

GitHub Copilot sets the stage, Cursor steals the show, LLMs fight back

GitHub Copilot kicked off the AI coding revolution with an interesting first take on code completion back in 2021, and by 2024, it had polished its act with enhanced chat interfaces and a broader model lineup. Initially powered by OpenAI’s GPT models, Copilot expanded its toolkit at GitHub Universe 2024, integrating Anthropic’s Claude 3.5 Sonnet, Google’s Gemini 1.5 Pro, and OpenAI’s o1 series. This multi-model flexibility let developers switch gears mid-task—Claude for refactoring, o1 for tricky edge cases—boosting productivity by up to 55%, according to GitHub’s own stats. It was a solid foundation, deeply embedded in Visual Studio Code and the GitHub ecosystem, but it was the team at Cursor who took the baton and ran with it, redefining the user experience in ways that left everyone else playing catch-up.

Cursor, a fork of VS Code, didn’t just tweak the formula—it rewrote it. By blending Claude 3.5 Sonnet and GPT-4o into a predictive autocomplete system with deep codebase awareness, Cursor turned coding into a vibe-driven flow. Type a hazy comment like “fetch user data,” and it doesn’t just suggest—it crafts a tailored, functional API call that fits your project’s DNA. The AI dev community took notice: Cursor’s adoption skyrocketed in 2024, with its revenue reportedly jumping from $1M to $100M in two years. Developers weren’t just using it—they were obsessed, dubbing it the gold standard for vibe coding. While Copilot laid the groundwork, Cursor’s intuitive UX and seamless integration made it the tool to beat, setting a pace that Replit, Cody, and even Copilot itself are now scrambling to match.

But the LLMs aren’t letting great IDEs like Cursor win the game that easily. Anthropic and xAI are firing back with Claude 3.7 Sonnet and Grok 3, released in February 2025, packing superior coding tools that threaten to outshine even the slickest interfaces. Claude 3.7 Sonnet, with its “extended thinking mode,” doesn’t just autocomplete—it reasons through multi-step problems, refining its logic on the fly. Its standout? Claude Code, a command-line tool that lets devs delegate entire programming tasks—think writing a full CRUD app from a spec—with a 15% jump in SWE-bench scores over Claude 3.5, hitting 63.7% unscaffolded and 70.3% with help. X posts in March 2025 buzzed with devs raving: “Claude 3.7’s CLI is like having a pair programmer who never sleeps.” Meanwhile, xAI’s Grok 3, trained on 10x the compute of Grok 2 via the Colossus supercomputer, brings “Big Brain” mode for complex debugging and a “DeepSearch” function that scours X and the web for real-time coding solutions. A Hacker News thread on March 20, 2025, lit up with praise: “Grok 3’s DeepSearch just saved me 3 hours on a gnarly regex bug.” These LLMs are flexing raw power—Cursor might own the UX, but Claude and Grok are proving that the models themselves can ship coding tools that rival, or even surpass, the IDEs built around them.

The collaborative edge with Replit, Cody, and V0

Replit took a different tack, aiming for an end-to-end solution that married coding, collaboration, and deployment. Its browser-based IDE, powered by Claude 3.5 Sonnet as of its 2024 Replit Agent release, let teams vibe code in real time—think Google Docs for developers. Need to prototype an app and push it live in under an hour? Replit’s got you covered, though its reliance on cloud services can nudge costs up for heavy users. Meanwhile, Cody, developed by Sourcegraph, mirrored Copilot’s strengths with its open-source roots and Code Graph API, offering context-aware suggestions that felt like having a senior dev whispering in your ear. Cody’s CEO, Beyang Liu, wants to industrialize coding, turning Cody and Sourcegraph into tools that let engineers on decade-old codebases with teams of thousands reclaim agility, using their skills to make systems hum instead of drowning in toil. They’re laser-focused on slashing the grunt work of sprawling legacy code, proving vibe coding isn’t just for shiny new projects—it’s for the grizzled pros too. Then there’s V0, Vercel’s sleek entry into the AI coding arena. Launched as a companion to its deployment platform, V0 turned heads in 2024 with its focus on rapid web development. Feed it a prompt or a Figma screenshot, and it’d churn out React or Next.js code with over 70% accuracy, per benchmarks from early 2025. It’s not just about output—V0’s polished UX made it a favorite for developers who value speed and aesthetics, though its web-centric focus left non-frontend folks looking elsewhere.

The paradigm shift brings vibe coding

By 2025, software development wasn’t just about writing code—it was about vibing with it. Karpathy’s “vibe coding” concept describes this shift: coding by prompting, tweaking, and iterating in a fluid, almost musical rhythm with AI. New players like Bolt, Windsurf, and Loveable amplified this trend. Bolt, from Stackblitz, let you spin up full-stack apps from a single prompt, while Windsurf—another VS Code fork—wowed beginners and pros alike with its Cascade mode, boasting real-time codebase awareness that outshone even Cursor in some tests. Loveable, meanwhile, leaned into simplicity, making app-building feel like a creative jam session rather than a grind.

Claude 3.5 Sonnet emerged as the LLM of choice powering this wave, with its upgraded March 2025 release pushing SWE-bench scores to 49%, outpacing OpenAI’s o1-preview. Its knack for debugging nested functions and handling massive token contexts made it a perfect fit for vibe coding’s freeform style. Yet, raw LLM power was just the foundation—tools like Replit and Cursor wrapped it in workflows so intuitive that developers could focus on ideas, not syntax.

Facts and Futures

The numbers tell a wild story: Cursor’s revenue reportedly soared from $1M to $100M in annual recurring revenue in just two years, a neon sign of vibe coding’s magnetic pull. But the real fireworks are happening across the ecosystem. Replit, the cloud-based IDE juggernaut, boasts over 40 million users globally as of early 2025, fueled by its Replit Agent—an AI powerhouse that spins up full-stack apps from plain-English prompts. Want a social media dashboard or a game in your browser? Tell Replit Agent, and it’ll scaffold, code, and deploy it while you tweak it in real time with a team—or solo on your phone. Posts on X rave about its accessibility: “Runs in the cloud, can do it on your phone, doesn’t even show the code unless you ask,” one user noted. Replit’s founder, Amjad Masad, dropped a bombshell on X in February 2025: “75% of Replit customers never write a single line of code!!” That’s vibe coding at scale—millions hacking out apps without ever touching a semicolon.

Windsurf, from Codeium, is another rocket in the sky. Its Cascade mode—an agentic chatbot that edits multiple files with eerie context awareness—has devs buzzing. “Windsurf’s Cascade blows other tools out of the water for evolving codebases,” a Reddit thread cheered in late 2024. Launched in 2024, Windsurf hit the ground running, with benchmarks by March 2025 showing it leading in prompt-to-API tasks, outpacing even Cursor in some tests. Meanwhile, Loveable’s chat-driven interface is a vibe coder’s dream: describe your app, and it builds a gorgeous frontend—complete with Supabase databases or Stripe payments—in hours. It’s racked up $10M in revenue just two months after monetizing, per Andreessen Horowitz’s February 2025 deep dive. And V0, Vercel’s sleek weapon, nails web development with a 90% success rate turning Figma designs into React code, per AIMultiple’s latest tests. “I built my professional site and a text-corrector app with V0,” an X user boasted in March 2025—testament to its speed and polish.

Here’s the kicker: the dev community isn’t just getting new toys—they’re getting better ones, fast. Tools like Windsurf’s real-time diff tracking, Replit’s zero-setup deployment, and Loveable’s one-click GitHub sync are sharpening the craft, letting seasoned coders vibe at lightspeed. But the real revolution? The flood of new users who’ve never coded before. They’re not devs—they’re vibe coders, pure and simple. Non-technical founders, designers, even kids are hacking out apps with these tools, from to-do lists to e-commerce sites. Replit’s 40 million users aren’t all pros; many are first-timers who heard Karpathy’s call to “give in to the vibes” and ran with it. Lovable’s pitch—“build software with a chat”—has non-coders launching MVPs in a day. V0’s Figma-to-code magic has designers shipping live sites without a dev in sight. Windsurf’s polished UX? Beginners are bootstrapping projects in 20 minutes.

This isn’t just a trend—it’s a tectonic shift. The dev community’s toolkit is exploding with power, but it’s also cracking the door wide open for anyone with an idea. Vibe coding’s promise, as Karpathy foresaw, is that creativity trumps syntax. In 2025, these tools aren’t just amplifying coders—they’re birthing a new breed of hackers who vibe their way to apps that work, ship, and sometimes even sell. What’s next? As Replit scales its AI agents, Windsurf refines its flows, Loveable nails integrations, and V0 perfects frontend finesse, software development might ditch the keyboard entirely. Grab your IDE—or just your voice—and let the vibe take over.

Top 5 AI coding apps that have vibes