TL;DR: Vibe coding works best when your AI has context, tools and guardrails. Below are real plugins you can install today, grouped by job, plus the steps to set them up. No sponsors, no favourites. Pick what fits your stack.
🧭 Step 1: Pick your AI coding home base
| Type | Examples | Good for |
|---|---|---|
| AI-first editors | Cursor, Windsurf, Zed | Deep multi-file edits and agent mode |
| Terminal agents | Claude Code, Aider, Gemini CLI, OpenAI Codex CLI | Git-friendly work from the command line |
| Plugins for VS Code / JetBrains | See Step 2 👇 | Keeping the editor you already know |
💡 Try two for a week on the same small project. Keep the one that understands your code with the least nudging.
🤖 Step 2: AI assistant plugins (VS Code & JetBrains)
| Plugin | What it does | Cost |
|---|---|---|
| GitHub Copilot | Autocomplete, chat and agent mode inside VS Code, JetBrains, Visual Studio | Free tier + paid |
| Continue | Open-source chat and autocomplete; plug in any model, including local ones | Free, bring your own model |
| Cline | Open-source agent that plans, edits files and runs terminal commands, with approval steps | Free, bring your own model |
| Roo Code | A Cline fork with custom "modes" (architect, coder, debugger) | Free, bring your own model |
| Gemini Code Assist | Google's assistant for VS Code and JetBrains | Free tier + paid |
| Amazon Q Developer | AWS-aware coding help and security scans | Free tier + paid |
| Tabnine | Autocomplete with privacy-focused and self-hosted options | Paid |
| JetBrains AI Assistant | Built-in AI for IntelliJ, PyCharm, WebStorm and others | Free tier + paid |
🔌 Step 3: Give your AI real tools with MCP servers
The Model Context Protocol (MCP) is an open standard that lets an AI assistant call outside tools. Cursor, VS Code (Copilot agent mode), Claude Code, Cline, Windsurf and others support it.
| MCP server | Why it helps |
|---|---|
| GitHub MCP Server (official) | Read issues, create branches and PRs, review diffs |
| Playwright MCP (Microsoft) | AI opens your app in a real browser, clicks around and checks its own work |
| Chrome DevTools MCP (Google) | Reads console errors, network calls and performance traces |
| Context7 (Upstash) | Pulls current, version-specific library docs, so fewer made-up APIs |
| Supabase / Postgres / MongoDB MCP | AI sees your real database schema instead of guessing |
| Sentry MCP | AI reads real production errors and stack traces |
| Figma Dev Mode MCP | Turns real design data into code that matches the design |
| Filesystem & Memory (reference servers) | Safe file access and long-term notes between chats |
⚠️ Only connect what the task needs, and start read-only where you can. Every tool is more access.
🧩 Step 4: Plugin bundles & skills for agents
Some editors now have plugin marketplaces that bundle an MCP server with ready-made skills (step-by-step recipes the AI follows). Examples in the Cursor marketplace:
- Superpowers: a skills library for test-driven development, debugging and planning.
- Vercel / Netlify / Cloudflare / Railway: deploy and debug from chat.
- Stripe / PayPal: payment integration guidance plus API access.
- Postman: mock, test and document APIs.
- SonarQube: code quality and secrets checks inside the agent loop.
- Linear / Notion: pull tasks and specs straight into the chat.
Write your own too. Each time you explain the same thing twice ("add a new API route", "write Playwright tests for this page"), save it as a skill or command.
📝 Step 5: Teach your AI your project (rules files)
AI forgets everything between chats. Add a short rules file to your repo:
-
AGENTS.md(a growing shared convention),.cursor/rules/,CLAUDE.md,.clinerules, or.github/copilot-instructions.md, depending on your tool.
# Project rules - Stack: Next.js, TypeScript strict, Tailwind - Run `npm test` before saying a task is done - Never edit files in /generated - No new dependencies without asking 🛡️ Step 6: Guardrail plugins so "vibes" don't break prod
| Plugin | Job |
|---|---|
| ESLint / Prettier / Biome | Lint and format; let the AI fix its own warnings |
| Error Lens | Shows errors inline so you (and the AI) spot them fast |
| Playwright Test for VS Code | Run and record end-to-end tests from the editor |
| Vitest / Jest extensions | Run unit tests next to the code |
| SonarQube for IDE (was SonarLint) | Catches bugs and security issues as you type |
| Snyk Security | Flags vulnerable code and packages |
| GitLens | See who changed what and why; easy rollbacks |
And always: small commits, and read every diff like a teammate's PR.
🚀 Step 7: Automate the boring loop
- CI: GitHub Actions or GitLab CI run tests on every push.
- AI code review: CodeRabbit, Greptile or Copilot code review add a second pair of eyes on PRs.
- Background agents: several tools can take an issue and open a PR on their own. Great for small, clear tasks.
- API testing: Thunder Client or REST Client in VS Code for quick endpoint checks.
🧠 Step 8: Prompting habits that matter more than any plugin
- ✅ Be specific: "Add email validation to
signup.tswith zod and 3 tests" beats "fix the form." - ✅ Plan first: ask for a plan, approve it, then let it code.
- ✅ Paste the full error, not "it doesn't work."
- ✅ One task per chat for cleaner context.
- ✅ Ask "why" so you learn and catch mistakes.
💸 Free starter stack
| Need | Free option |
|---|---|
| Editor + AI | VS Code + Continue or Cline, or a free tier of Copilot |
| Local models | Ollama or LM Studio |
| MCP | GitHub MCP + Playwright MCP + Context7 |
| Testing | Playwright, Vitest |
| Security | SonarQube for IDE, Dependabot |
| Hosting | Free tiers on Vercel, Netlify or Cloudflare |
Free tiers change often. Check each site before you commit.
✅ Quick checklist
- [ ] Picked an AI editor, plugin or terminal agent
- [ ] Added 2–3 MCP servers I actually need
- [ ] Added a project rules file
- [ ] Linter, tests and a security plugin running
- [ ] Reviewing every diff before merging
🎤 Your turn
There's no single "best" setup. It depends on your stack, budget and how much control you want. What's in your AI dev toolkit? Drop your favourite plugin or MCP server in the comments 👇
