What vibe coding is, when it works vs fails, security risks from real studies, and how MCP connects vibe-coded apps to live crypto data through DexPaprika.

Vibe coding: build with AI agents instead of writing every line

Vibe coding is a software development practice where you describe what you want in natural language and let an AI write the code, accepting the result without reading or fully understanding it. Coined by Andrej Karpathy in February 2025, the term captures a loop: describe, generate, accept, run, paste errors back in.

Within a year, "vibe coding" went from a throwaway tweet to Collins Dictionary's Word of the Year for 2025. By the 2025 Stack Overflow survey, 84% of developers were using AI tools in their workflow and 51% used them daily. This isn't a niche practice anymore.

Karpathy's original description is still the best one: "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works." You talk to the AI. It writes code. You run it. If it breaks, you paste the error back in. You never open the file to read what it actually wrote. The code "grows beyond my usual comprehension," as he put it, and that's fine because it's a weekend project.

This is fundamentally different from AI-assisted coding, where a developer uses Copilot suggestions but still reviews every line. Vibe coding means giving up that control entirely. You're the product director, not the engineer. Simon Willison drew the clearest line: "Not all AI-assisted programming is vibe coding." The distinguishing feature is accepting code you don't understand.

How vibe coding works

The workflow is a five-step loop that keeps repeating until you're done:

1. Describe what you want ("build a dashboard showing top Ethereum pools by volume")
2. Generate the AI writes the full code
3. Accept All without reading the diff
4. Run the app and observe what happens
5. Paste errors back if something breaks, with no additional context

No architecture planning, no code review, no manual debugging. Karpathy used Cursor with Claude Sonnet and SuperWhisper for voice input, so he barely touched the keyboard. In practice, a session looks like this: you say "add a login form to the sidebar," the AI rewrites three files, you click Accept All, refresh the browser, see the form, and say "make the button blue." Fifteen exchanges later you have a working app and no idea what's in the codebase.

The tools that make this possible have exploded in the past year. Cursor hit $2 billion+ in annualized revenue by March 2026, the fastest B2B SaaS to $1 billion ever. GitHub Copilot has 20 million users and 42% market share among paid AI coding tools. Claude Code runs as a CLI agent that reads your codebase, writes code, runs tests, and even submits PRs autonomously. It scored 72.5% on SWE-bench, which measures real-world software engineering tasks.

For people who aren't developers at all, tools like Lovable ($100 million ARR in 8 months), Bolt.new, Replit Agent, and v0 generate entire applications from a prompt. 25% of Y Combinator's Winter 2025 batch had codebases that were 95%+ AI-generated. Microsoft CEO Satya Nadella said up to 30% of Microsoft's own code is now AI-written. These numbers were unthinkable two years ago.

When vibe coding works (and when it doesn't)

I want to be direct about this because the hype consistently gets ahead of the evidence.

Vibe coding is genuinely great for prototypes, weekend projects, internal tools, and MVPs. Pieter Levels built a multiplayer flight simulator with Cursor and Grok 3, hit $1 million ARR within 17 days. He's an experienced developer, but the speed was pure vibe coding: prompt, accept, ship. Non-developer Paulius Masalskas used the same approach to build CreatorHunter, a creator analytics tool that generated $30,000 in revenue. These aren't toy demos. They're real products making real money from people who couldn't have built them this fast before.

But production systems are a different story. The research is stacking up, and it's not all good news:

The METR finding is the most counterintuitive. Those developers thought they were 24% faster. They were actually 19% slower. The perception gap alone should give you pause.

Karpathy himself backed away from pure vibe coding by late 2025, acknowledging the need for "more oversight and scrutiny." The consensus I've seen settle is: vibe code to explore, then switch to disciplined AI-assisted development for anything that faces real users.

Vibe coding with live data: how MCP connects everything

Here's what I think separates vibe-coded toys from vibe-coded tools that actually do something: connecting them to live data through MCP.

Without MCP, a vibe-coded app gets mock data or hardcoded values. You end up with a beautiful dashboard showing fake numbers. With MCP, you say "show me the top 10 Ethereum pools by volume" and the AI pulls current data because it has a direct connection to the source. The app works with actual market data from the first run.

I set up DexPaprika's MCP server in Cursor to test this. Two lines of config:

// .cursor/mcp.json
{
  "mcpServers": {
    "dexpaprika": {
      "url": "https://mcp.dexpaprika.com/streamable-http"
    }
  }
}

Then I started prompting:

• "Build me a dashboard showing the top 10 Ethereum DEX pools sorted by 24h volume" — the AI picked the getNetworkPools tool, passed order_by: volume_usd, and generated a working React table with real pool data.
• "Add a price tracker for WETH, USDC, and SOL with auto-refresh" — it used getTokenDetails for each, set up a polling interval, and displayed live prices.
• "Show me pools created in the last 24 hours with more than $100K volume" — it chained getNetworkPoolsFilter with the right parameters and rendered a sortable list.

The AI reads the MCP server's 14 typed tools, picks the right one, and generates working code that calls the API correctly. No API key. No registration. No documentation to read. That zero-friction model matches the vibe coding philosophy perfectly.

For Claude Code, the same data is available via the marketplace plugin. For centralized exchange data, CoinPaprika's MCP server adds 23 free tools covering 2,500+ cryptocurrencies. DexPaprika also publishes skill files that teach the AI about endpoint quirks and field naming conventions, so it doesn't trip over the API. The agents.dexpaprika.com hub ties all the integration paths together, and the docs include a dedicated vibe coding tutorial.

If you want to go deeper and build your own MCP server to connect custom data, our guide to building your first MCP server walks through the process.

Vibe coding: frequently asked questions

What to remember about vibe coding