Best vibe coding tools
By Mari Luukkainen · 10 min read · Last updated
On this page
- What kinds of vibe coding tools exist?
- Prices and limits (checked 1 November 2026)
- What changed in 2026
- Browser builders
- AI editors
- Terminal agents
- Which tool should you pick for your first project?
- What matters more than the tool
- What does the same task look like in each group?
- How do you test a tool in one hour?
- Frequently asked questions
The best vibe coding tool for a first project is a browser builder such as Lovable, Bolt.new or Replit, because nothing needs installing and all three have a free plan. For code you want to own and edit, use an AI editor such as Cursor. For delegating whole tasks, use a terminal agent such as Claude Code or Codex.
This page sorts the tools by how you work with them, then gives prices and free tiers checked on 1 November 2026, a pick, and the tools whose status has changed this year. New to the idea? Read what vibe coding is first.
Ask AI about this pageChatGPTClaudePerplexityGoogle AI ModeGrok
What kinds of vibe coding tools exist?
The tools split into three groups. The group decides where your project lives, how much of the code you see and how much the AI does without asking.
| Group | Examples | Where the project lives | You see the code | Best for |
|---|---|---|---|---|
| Browser builder | Lovable, Bolt.new, Replit, v0, Macaly | In the vendor's cloud, exportable | Beside the preview, optional | A first project with no set-up |
| AI editor | Cursor, GitHub Copilot in VS Code | A folder on your computer | Always, in the editor | Learning the code while building |
| Terminal agent | Claude Code, Codex, OpenCode, Antigravity CLI | A folder on your computer | On request, as a diff | Larger changes with less supervision |
The groups overlap. Cursor runs terminal commands in its agent mode, has cloud agents and a command line tool. Claude Code has a VS Code extension and a desktop app. Codex runs in a terminal, an editor extension, the ChatGPT apps and in the cloud.
Prices and limits (checked 1 November 2026)
Prices are monthly in US dollars as the vendors list them. Plans and limits change often, and several changed in 2026. Each row links to the vendor page I read. For Claude Code limits, API costs and how to choose a plan, see Claude Code pricing.
| Tool | Free tier | Paid plans | Source |
|---|---|---|---|
| Claude Code (Claude plans) | The free Claude plan does not include Claude Code | Pro $20 ($17 billed annually), Max $100 (5x Pro usage), Ultra $200 (10x Pro weekly usage, no 5-hour limit) | claude.com/pricing, setup docs |
| Cursor | Free: limited agent and chat requests | Plus $8 (Grok models only), Premium $30 (4x Plus usage, cloud agents), Super $100, Ultra $200. Other vendors' models such as Claude are billed by usage | cursor.com/pricing |
| GitHub Copilot | Free: 2,000 completions a month, limited chat and agent use | Pro $10 (1,500 AI credits), Pro+ $39 (7,000 AI credits). One credit is $0.01 | GitHub plans |
| Codex (ChatGPT plans) | Included in ChatGPT Free | Go $8, Plus $20, Pro from $100, Business $20 per user billed annually | ChatGPT pricing docs |
| Antigravity CLI (Google) | Individuals $0 with basic weekly rate limits | Higher limits with Google AI Pro or Ultra, business from $30 per seat | antigravity.google/pricing |
| OpenCode | Free and open source (MIT). Some models free for a limited time | OpenCode Models is pay as you go with a $20 minimum top-up. OpenCode Go $10 | opencode.ai, Models, Go |
| Devin Desktop (formerly Windsurf) | Free: light quota | Pro $20, Max $200, Teams $80 plus $40 per developer seat | devin.ai/pricing |
| Kiro | Free: 50 credits a month | Pro $20 (1,000 credits), Pro+ $40, Pro Max $100, Power $200 | kiro.dev/pricing |
| Amp | Hobby: free, bring your own keys | Individual $20 | ampcode.com/pricing |
| Jules (Google) | Base: 15 tasks a day, 3 at a time | 100 tasks a day with Google AI Pro, 300 with Google AI Ultra | jules.google |
| Mistral Vibe | Free: limited coding sessions | Pro $14.99, Team $24.99 per user | mistral.ai/pricing |
| Lovable | 5 build credits a day, 30 a month at most | Pro from $25 (100 credits), Business from $50 | Lovable plans |
| Bolt.new | 1M tokens a month, 300K a day | Pro $25 (10M tokens), Teams $30 per member | bolt.new/pricing |
| Replit | Starter: 1 free published app, link goes down after 30 days | Core $20 ($18 annual), Pro $100 ($90 annual) | replit.com/pricing |
| Macaly | 1 published app on a Macaly domain, 1M credits a month | Starter $10 (custom domain), Pro $30 (5 apps, 10M credits), Max $90 (15 apps, 30M credits) | macaly.com/pricing, checked 12 November 2026 |
| v0 | 7 messages a day | Plus $30 per user, Business $100 per user | v0.app/pricing |
Two prices I could not confirm: whether Google AI Studio's Build mode is free (Google lists no price), and monthly-billed prices for Base44 and Bubble (only annual prices are shown, see best AI app builders).
What changed in 2026
- Cursor is owned by SpaceX. SpaceX agreed to buy it for $60 billion in stock in June 2026 (Engadget), and Cursor announced the close on 14 August 2026 (GIGAZINE). The Hobby, Pro and Pro+ plan names are gone from the pricing page, so older $20 Pro figures are outdated.
- Gemini CLI has no free tier since 18 June 2026. Free individual access moved to Antigravity CLI, and Gemini CLI is now for Gemini Code Assist Standard and Enterprise customers (Google).
- Windsurf is now Devin Desktop, from Cognition (announcement, 2 June 2026).
- Firebase Studio takes no new workspaces and shuts down on 22 March 2027 (firebase.studio).
- Mistral's Le Chat is now part of Vibe.
- Claude has a new $200 Ultra tier, and Max is a single $100 tier.
Browser builders
You type into a chat and a preview of the app appears beside it. The vendor runs the project, so there is no installation and no hosting set-up for the first version. I have never used a browser builder, so what follows comes from the vendors' own documentation.
- Lovable. Chat-first builder for web apps. Lovable Cloud adds a built-in backend with a database, storage, authentication and functions, built on Supabase's open source foundation (docs). Check Row Level Security on every table, because an insufficient policy is what CVE-2025-48757 describes.
- Bolt.new. Generates and runs a web app from a prompt.
- Replit. An online development environment with an AI agent that can host what you build. In July 2025 its agent deleted a production database during a SaaStr test (Fortune). Keep real data out of anything you let an agent run unattended.
- Macaly. A builder for websites and web apps from Macaly, a company from the Czech Republic. Hosting, a database and sign-up and login are included on every plan, and the Free plan needs no card (pricing). On Free, the backend pauses when you reach the allowance until you upgrade. I could not confirm from its site whether you can export the code.
- v0. Vercel's builder. It started as interface components and now builds "from landing pages, dashboards, ecommerce stores, AI apps, to full-stack applications" with one-click deploy on Vercel (docs).
The weak spot is control. When the builder loops, you have few levers besides rewording the prompt. Past a certain size, most people export the code and continue in an editor. More builders are in best AI app builders.
AI editors
An editor shows your files, and the AI works inside it. You chat about the project, the AI edits the files, and you see each change.
- Cursor. An editor built around AI. Its agent runs terminal commands and edits code, cloud agents (formerly background agents) produce merge-ready pull requests, and a command line tool called agent exists. It is the tool I used for my first vibe coded sites in 2025, including a Next.js migration where one prompt, "SEO the shit out of it", produced titles, meta descriptions, Open Graph tags and alt text (post). See Cursor vs Claude Code.
- GitHub Copilot. An assistant in VS Code and JetBrains products with completions, chat, agent mode, a command line tool and a model picker that includes Claude models. See Copilot vs Claude Code.
- Devin Desktop. The former Windsurf, an AI editor from Cognition.
Editors suit people who want to see what the code does. The cost is attention, because you review as you go.
Terminal agents
You open a terminal in your project folder and give the agent a task. It reads files, edits them, runs commands such as the tests, and reports what it did. You review the result.
- Claude Code. Anthropic's agent. It runs in the terminal, VS Code and JetBrains, a desktop app, on the web (cloud sessions that push branches and pull requests), in Slack and in GitHub Actions. Start with what Claude Code is and how to use it.
- Codex. OpenAI's agent. It runs in a CLI, an IDE extension, the ChatGPT apps and in the cloud. Compare it in Codex vs Claude Code.
- OpenCode. Open source, works with models from many providers, has a desktop app. See OpenCode vs Claude Code.
- Antigravity CLI and Gemini CLI. Google's terminal agents. See Gemini CLI vs Claude Code.
- Others. Amp, Kiro, Jules (asynchronous tasks) and Mistral Vibe CLI. See Claude Code alternatives.
Agents suit larger changes such as "add a settings page with these four fields, a test, and update the navigation". They run commands, so read what they ask to run before you approve it.
Which tool should you pick for your first project?
My pick comes from the criteria below and the prices above, not from having used every tool. I used Cursor for my first sites in March 2025 and have never used a browser builder. My March 2026 guide to building a marketing site names Cursor or Claude Code as the AI to use (post), and my 2026 posts describe handing failures to Claude.
| If this describes you | Start with | Move on when |
|---|---|---|
| You have never written code and want a result today | A browser builder on its free plan | You need a change the builder cannot make |
| You want to understand the code as it appears | An AI editor: Cursor Free or Copilot Free | You approve every edit without reading it |
| You will pay anyway and want to delegate whole tasks | A terminal agent: Claude Code (Pro, $20) or Codex (included in ChatGPT Plus, $20) | You cannot review the diffs fast enough |
| You want no lock-in to one vendor | OpenCode with your own API key | The set-up costs more time than it saves |
| You want a free terminal agent | Antigravity CLI (Individuals, $0) | The weekly rate limits stop your work |
Try two tools on the same small task in a throwaway project. The worked first project works as that task.
What matters more than the tool
- Project instructions. AGENTS.md is read by Codex, Claude Code, GitHub Copilot, OpenCode and Cursor. My own handbook is a folder of markdown files the AI reads before it starts, and it changed how Claude works in my repos (post).
- Small prompts with a clear goal. See the prompt library.
- Version control. Commit after every working round.
- Testing as a user. The AI cannot tell whether the app does what you meant.
What does the same task look like in each group?
Take one task: "Add a dark mode toggle that remembers the visitor's choice."
| Group | What you do | What to watch |
|---|---|---|
| Browser builder | Type the sentence in the chat and wait for the preview | Does the choice survive a refresh? Does the page flash white on load? |
| AI editor | Open the layout file, select it, ask in the chat panel, accept the edits | Did it edit only the global styles and the layout file? |
| Terminal agent | Write the goal and the check, approve the plan | Did it add a dependency for something you can do in ten lines? |
Add a dark mode toggle in the page header. The choice is saved in the browser and applied before the page is shown, so there is no white flash on reload. If the visitor has not chosen, follow the device setting. Do not add dependencies. Tell me which files you changed."Applied before the page is shown" is the detail a tool will not add unless you ask.
How do you test a tool in one hour?
This is a test you run yourself. I have not run it on these tools.
| Minute | Do | Record |
|---|---|---|
| 0 to 10 | Install or sign up, start a throwaway project | Number of steps you had to look up |
| 10 to 30 | Run the dark mode prompt | Interventions and wrong turns |
| 30 to 45 | Ask for a second, unrelated change | Did the first change survive? |
| 45 to 60 | Break something on purpose (rename a file) and ask it to fix the error | Did it find the cause or guess? |
Do the hour in two tools and compare notes.
Frequently asked questions
Who owns Cursor now?
SpaceX. It agreed to buy Cursor for $60 billion in stock in June 2026, and Cursor announced the acquisition had closed on 14 August 2026. Cursor is sold as part of The X Subscription with Grok.
What happened to Windsurf?
Windsurf is now Devin Desktop, from Cognition. Cognition announced the change on 2 June 2026, and windsurf.com now redirects to devin.ai/desktop.
Is Firebase Studio still available?
It takes no new workspaces and is being shut down on 22 March 2027, according to firebase.studio.
Can I use my Claude subscription in tools other than Claude Code?
Not in third-party tools. Anthropic does not permit third-party developers to route requests through Free, Pro, Max or Ultra plan credentials. Use an API key instead, or use Claude models inside Cursor or GitHub Copilot, which bill separately.
Is there a free terminal agent to try?
Yes, several. Antigravity CLI is $0 for individuals with basic weekly rate limits, OpenCode is free software and has some free models for a limited time, Codex is included in ChatGPT Free, and Amp's Hobby plan is free if you bring your own API keys.
Next guideBest AI app builders →