Vibe Coding

What is Vibe Coding?

Vibe coding is a way of programming where you describe what you want in plain language and let an AI agent do the typing. You stay in the loop reviewing diffs, running things, and steering, but you stop hand-writing every line.

The term was coined by Andrej Karpathy in this tweet:

Andrej Karpathy

@karpathy

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. Also I just talk to Composer with SuperWhisper

The free-tier landscape changed a lot in 2025 and 2026. Editor subscriptions used to be the obvious choice; today the best free setup is an open-source terminal agent plus a free or cheap model. Here's the setup, then a tour of the alternatives.

Best Free Setup: OpenCode

Start with OpenCode an open-source, MIT-licensed terminal agent built by the SST team. It's provider-agnostic, supports 75+ models including local ones via Ollama, and has a polished TUI plus a desktop app and IDE extensions for VS Code, Cursor, Zed, and JetBrains.

Install (pick one):

# macOS / Linux
curl -fsSL https://opencode.ai/install | bash

# Or via npm
npm i -g opencode-ai@latest

# Then run
opencode

Once installed, run the /connect command and sign in with OpenCode Zen OpenCode Zen is a curated set of models served by the OpenCode team. Some are free for limited periods. The current free lineup rotates as feedback windows close, so check the Zen page for what's available right now. If you want a stable cost-effective baseline, Anthropic's Claude (with a Pro/Max plan) and DeepSeek are the usual go-tos.

To help the agent see your codebase efficiently, pair it with Code Collector a VS Code extension that walks import graphs and packages just the relevant files into your prompt. Useful when you're switching to a model with a smaller context window.

Want a single endpoint for every model instead of juggling provider keys? UnoRouter routes across 200+ models from 35+ providers with automatic failover.

Choosing a Model

Different models are better at different things. As a rough 2026 starting point: Claude Sonnet/Opus for dense refactors and reasoning, Gemini 2.5 Pro when you need a giant context window, GPT-5 codex variants for general coding, DeepSeek V3 when you want strong quality at near-zero cost, and OpenCode Zen's currently-free models when you're experimenting.

For an independent ranking that updates as new models drop, check the LM Arena Web Development Leaderboard.

Don't lock in. Switching models in OpenCode is a single command, so try a few on the same task and keep whichever one wins for your stack.

Terminal Agents (Alternatives to OpenCode)

Other terminal agents worth knowing. The list is ordered roughly by popularity in 2026.

OpenCode

Open source. Provider-agnostic, with 75+ models and a free Zen tier that rotates.

Claude Code

Anthropic's official agent. Top-tier reasoning, especially with a Pro/Max subscription.

Gemini CLI

Google's official agent. 1,000 free requests/day with Gemini's 1M-token context.

Codex CLI

OpenAI's official agent. Pairs with ChatGPT Plus or pay-per-token via the API.

Aider

Git-native pair programmer. Architect/editor split keeps cost low across any provider.

Cline

Open-source VS Code extension. Bring your own key, including free providers.

AI-First Editors

If you'd rather have AI baked into a graphical editor than driven from a terminal, these are the modern picks.

Cursor

VS Code fork. Best supervised IDE experience, with multi-file Composer and parallel agents.

Windsurf

AI-native IDE from the Windsurf team. Cascade agent and free tier with daily message limits.

Antigravity

Google's AI-first coding environment. Plans, browses, and edits across multiple files.