Vibe Coder’s Guide to MVP
Resources · Craft & Compliance

Dev tools & agents

Claude Code vs Codex vs Cursor, the editors and CI worth running, and the toolchain we use to ship MVPs in days.

The single biggest productivity decision an MVP builder makes in 2026 is which AI coding agent to live in. Everything else — IDE, linter, CI — is downstream of that pick, because the agent decides how it wants to work and a good toolchain stays out of its way. This page covers the real options, the tradeoffs, and what we actually use.

Anthropic's official CLI agent. Runs in a terminal, plans long sequences of work, asks clarifying questions, and handles multi-file refactors without losing the thread. The 1M-context Opus tier is the model of choice for a single-shot brownfield change across a real codebase; Sonnet is the daily driver for normal work.

Why we recommend it for non-engineers specifically:

  • It explains what it's doing in conversational English by default. You can read the output and learn.
  • Long-horizon planning — it'll happily spend 20 minutes on one task without losing the goal.
  • Skills (.claude/skills/*.md) let you teach it project-specific patterns once and reuse them across sessions.
  • Subscription pricing ($20/mo Pro, $100/mo Max, $200/mo Max 20x) is predictable; no per-token surprise bills.

This site was built with Claude Code on AWS. The full comparison with Codex is in /blog/claude-code-vs-codex-comparison/.

Codex (OpenAI)

OpenAI's coding agent, available as a CLI and a web/IDE surface. Built on the GPT-5 family. Faster turn times than Claude Code on focused tasks, tighter output, less conversational. Excellent when you already know exactly what you want and want it executed quickly.

Pricing follows the ChatGPT subscription tiers (Plus $20/mo, Pro $200/mo) plus optional API metered usage for power users. The skill bundle's .codex/skills/ convention works the same way as .claude/skills/, so the same skills work in both agents.

If you're already a comfortable developer and value speed over hand-holding, Codex is the pick.

Cursor

A VS Code fork with deep AI integration baked into the editor. $20/mo Pro tier in 2026, $40/mo Business. The tab-completion model (Composer) is the best in class for the "I'm typing and the agent finishes my thought" workflow, and the agent panel handles multi-file edits without leaving the editor. The tradeoff: you're in an IDE, not a terminal, which some people love and some hate.

Cursor is the right pick if you genuinely want to be in the code yourself most of the time and have an agent on standby. It's the wrong pick if you want the agent to drive and you to review.

Aider

Open-source CLI agent, BYOK (you bring your own Anthropic / OpenAI / etc. API key). Mature, fast, terminal-native, integrates with git so every change is a real commit. The model selection is yours; pricing is whatever your API bill is.

Aider is the choice for developers who want full control over the agent's prompt, model, and cost, and who are comfortable in a terminal. Less polished than Claude Code or Codex, more configurable.

Continue

Open-source extension for VS Code and JetBrains IDEs. BYOK. Think of it as the "Cursor experience" but as a plugin you can install in your existing editor. Smaller team than Cursor, more flexible model routing.

Reasonable pick if you already have a VS Code or JetBrains setup you love and you want AI in it without switching editors.

GitHub Copilot

Microsoft's autocomplete-style assistant. $10/mo individual, $19/mo business in 2026. The agent mode added in 2024 closed some of the gap with Cursor, but the core experience is still inline completions, not agentic work.

Copilot is fine as a typing assistant. It is not the right tool if you want the agent to plan, edit multiple files, run tests, and iterate. Don't pick Copilot expecting Claude Code.

Agent comparison

Agent Pricing (2026) Style IDE-integrated BYOK Free tier
Claude Code $20–$200/mo (sub) Agentic, conversational No (CLI) Optional API Limited via Claude.ai
Codex $20–$200/mo (sub) Agentic, terse Yes (web + CLI) Optional API Limited via ChatGPT
Cursor $20–$40/mo Hybrid editor + agent Yes (own IDE) Optional Free tier (limited)
Aider API cost only Agentic, terminal No (CLI) Required Free if you use free models
Continue API cost only Hybrid plugin Yes Required Free if you use free models
Copilot $10–$19/mo Autocomplete + light agent Yes No Free for students/OSS

VS Code

Microsoft's editor. Free, open-source-ish, the dominant choice. The extension ecosystem is unmatched. If you don't have strong feelings about editors, install VS Code, install the relevant agent extension or just run Claude Code in the integrated terminal, and move on.

Cursor (as an editor)

Same product as the agent above. As an editor it's a VS Code fork, so anything that works in VS Code works in Cursor. The premium features are the AI ones.

Zed

Newer Rust-based editor from the team that built Atom. Very fast, clean UI, native multiplayer collaboration, growing AI integration. Free for individuals. Worth a look if VS Code feels heavy.

JetBrains IDEs (WebStorm, PyCharm, etc.)

Paid ($16+/mo individual or $69+/mo business per IDE, or All Products Pack ~$28/mo individual). Best-in-class refactoring, debugging, and language intelligence. The AI Assistant has caught up but isn't the headline reason to be here. Pick JetBrains if you'll spend most of your time reading and refactoring code yourself.

Vim / Neovim

Free, terminal-native, infinite customization. The learning curve is real and the AI integration story is fragmented (CodeCompanion, Copilot.vim, custom Aider integrations). For someone already deep in a Vim workflow, fine. For a non-engineer starting fresh, no.

TypeScript LSP

Built into VS Code and most editors. Free. The single most valuable piece of code intelligence on a JS/TS project — it gives you autocomplete, go-to-definition, type errors as you type. If you skip TypeScript on an MVP to "go faster," you're trading a week of LSP-aided speedup for a few hours of setup. Don't.

Biome / oxlint (modern fast linters)

Both written in Rust, both replacing ESLint for new projects. Biome is mature, includes formatting (replacing Prettier), runs ~10–25x faster than ESLint. Oxlint is newer, faster still, formatter is in progress. Either is a vast upgrade over the legacy ESLint + Prettier combo.

# add Biome to a fresh project
npm install --save-dev @biomejs/biome
npx biome init
npx biome check --write .

Prettier

The de-facto formatter for years. Still excellent. If you're already on Prettier and not chasing speed, no reason to switch. For a new project, Biome covers both linting and formatting in one tool.

GitHub

Default repo host. Free for unlimited public and private repos with reasonable Actions minutes (2,000/mo private). Pull requests, issues, the works. The only reason to look elsewhere is values-based or scale-based.

GitLab

Similar feature set to GitHub plus a built-in container registry, more generous CI minutes on the free tier. Slightly less polished UI. Self-hostable. Good pick if you want CI bundled or if you're going self-hosted later.

Codeberg

Non-profit, EU-based, Forgejo backend. Free for FOSS projects, no commercial use without arrangement. Pick this if values matter and your project is open source.

Sourcehut

Minimal, email-driven workflow (patches via email, no PRs). Power-user pick. Cheap, fast, opinionated. Not the right call for a team that hasn't done email-patches before.

CI/CD

Option Best for Free tier
GitHub Actions Default, in-repo workflows 2,000 min/mo private, unlimited public
Vercel deploys Frontends auto-deploying on push Generous; usage-based after
AWS CodeBuild Anyone already on AWS 100 build-min/mo
GitLab CI Anyone on GitLab 400 min/mo on free tier
CircleCI Legacy choice; still works 6,000 build-min/mo

For a Vite + React MVP, the simplest pipeline is: push to GitHub, Vercel (or your S3 bucket via a GitHub Action) deploys. Nothing fancier needed until you have a real backend.

iTerm2 (Mac)

Free, mature, what most Mac developers run. Tabs, splits, profiles, the works. Excellent default.

Warp

Modern terminal with AI built in. Free for individuals, paid for teams. The AI features (command suggestions, error explanations) are useful but if you're already running an agent in the terminal, you're double-paying for AI.

Wezterm

Cross-platform (Mac, Linux, Windows), GPU-accelerated, configured in Lua. Fast and keyboard-driven. Free.

Ghostty

New terminal from the founder of HashiCorp. Fast, minimal, native rendering. Free, open source. Worth a look if iTerm2 feels dated.

Skill bundles

The Vibe Coder's Guide skills are markdown files that teach an agent project-specific patterns: design rules, compliance steps, deploy sequences, the works. They live in .claude/skills/*.md for Claude Code and .codex/skills/*.md for Codex, and the conventions are interchangeable — the same markdown works in both agents because both are reading text.

A typical skill is 200–500 lines of markdown describing a workflow ("write a privacy policy by walking the user through these 9 questions, then output to /privacy/index.md"). The agent loads it on demand, follows the steps, and you get repeatable output instead of re-prompting from scratch every time. We covered the philosophy in /blog/skills-not-prompts/.

Our recommendation

Claude Code if you're a non-engineer. Longer planning horizon, more handholding, easier to learn from. Pair it with VS Code for the rare moments you want to read code yourself, and run it from the integrated terminal.

Codex if you're already a developer. Faster turn times, more terse, gets out of your way.

Both work with the Vibe MVP skills. You don't have to pick irrevocably — try each for 30 minutes on the same task and see which one matches how you think. The skill files port between them.

For everything else: GitHub for the repo, GitHub Actions for CI, Vercel or S3+CloudFront for deploy, Biome for lint+format, Lucide for icons (covered in /resources/design-ui/), iTerm2 or Ghostty for the terminal.

Anti-patterns we keep seeing

  • Picking Copilot when you need agentic work. Copilot is autocomplete. If you want the agent to plan a feature, write it across five files, run tests, and iterate, you need Claude Code, Codex, or Cursor — not Copilot.
  • Picking the agent based on Twitter hype. Spend 30 minutes with each on the same real task. The right agent for you is the one whose output you can read without translating. Hype is not a substitute for fit.
  • Skipping TypeScript to "move faster." You'll spend the saved time debugging undefined-is-not-a-function errors at 2 AM.
  • Running ESLint + Prettier in 2026. Biome does both, faster. Migrate.
  • Three different terminals. Pick one. Configure it once. Stop fiddling.
  • Manual deploys. If git push doesn't deploy, you'll skip deploys. Wire CI on day one — it's a 10-minute job.
← Previous
Legal & compliance