Ride the wave with a crew of AI agents.
VibeCode.surf is a desktop code editor built around agents. Say what you want. Several agents plan it, write it, run it and test it in a real browser, on your own computer, with any model you choose.
curl -fsSL https://vibecode.surf/install.sh | shirm https://vibecode.surf/install.ps1 | iexcurl -fsSL https://vibecode.surf/install.sh | shOne command installs the desktop app. A free account keeps it up to date; your code and API keys stay on your machine. Watch the 75-second demo ↓
src/ui/Forecast.tsx.PIPELINE · THURSDAY
Best day this week
NEARBY SPOTS
Documents
Media
Swellcast
A surf forecast for the week ahead. Pick a spot, see the wave height each day, and know which day is worth the drive.
Run it
npm install npm run dev
How the best day is chosen
Each day gets a score from wave height, period and wind direction. The highest score wins; ties go to the earlier day.
A · claude-opusfinishedPIPELINE · THURSDAY
B · gemini-profinishedPipeline
See it before you install it.
Start with the 75-second demo. Most of the others are recorded from the real app, building a real project and then running it.
One agent is a helper. Several are a crew.
A lead agent splits the job across up to six helpers that work at the same time. Or add agents yourself and talk to each one in its own tab.
They share the wave without dropping in on each other
- Each helper owns its files and can write only there. Nobody else can write in them while it works.
- No blind overwrites: an agent cannot replace a file another agent changed since it last read it.
- One command at a time per project, so installs and builds never collide.
- A different model for each: a strong one on the lead, fast cheap ones on the helpers.
- Keeps going without you: close the window and the crew carries on. Quit, and they pick up where they stopped next time.
Pick a board for every wave. Any model, any provider.
Connect the providers you already pay for, or run models on your own computer for free. Each agent can ride a different one.
Claude
OpenAI
OpenRouter
Gemini
Groq
DeepSeek
Z.ai GLM
Mistral
xAI
Kimi
Qwen
Together AI
Fireworks
Cerebras
Cloudflare
LiteLLM
Ollama
LM StudioNot on the rack? Add it as Custom: any server that speaks the OpenAI format works, with its address and key. Model lists and prices come from models.dev.
Pictures, video and voice for your agents
Higgsfieldimages and video
fal.aiimages and video
Replicateimages and video
Runwarefast images
Ideogramimages with text
FLUXthrough fal.ai, Replicate, Together, Runware
Klingvideo, through fal.ai or Replicate
Seedancevideo, through fal.ai or Replicate
ElevenLabsvoices and transcription
Deepgramdictation and voices
AssemblyAItranscriptionAgents use these to make the images, clips, voice-overs and transcripts your project needs, and save them straight into it. All names are trademarks of their owners.
Choose the model, how hard it thinks, and see the bill
- Searchable model list grouped by provider, with each model's input and output price.
- Thinking level per agent, from Off to Max. The menu tells you what the selected model supports.
- Tokens and cost under every reply, a running total per project, and a breakdown per model.
- It recovers by itself when a connection drops, credit runs low, or a model cannot read images.
Two models, one request. Keep the winner.
Every model solves a problem its own way. Send the same request to two of them, look at both results side by side, and keep the one you like.
- Each model gets its own copy of your project. They work at the same time and cannot touch each other, or your real files.
- Judge by looking. Two live previews next to each other, with the changed files, the time and the cost of each.
- One click to keep it. The result you choose becomes your project, the other one disappears, and you can still undo it.
- Any pairing. Claude against Gemini, a local model against a hosted one, or the same model twice for a second opinion.
Describe it. Watch it take shape.
Web, mobile, desktop, games, tools, back ends. In whatever language the job needs: Python, Go, Rust, Swift, Kotlin, TypeScript and many more.
Web apps and sites
Landing pages, SaaS products, shops and internal tools, with a live preview beside the chat.
Mobile apps
Checked at phone and tablet size by the agents themselves before they call it done.
Dashboards and data
Drop in a file, ask a question, get charts and the code that made them.
Games and toys
Small ideas turned into something you can play in minutes.
1,847 words · 142 lines
surf 41 · wave 37 · tide 22
Scripts and command-line tools
Written, run and tested in the built-in terminal.
{ "height_ft": 6.4,
"period_s": 14 }
APIs and back ends
Routes, database, tests and a README, in the stack you ask for.
Write it. Run it. Look at it. Fix it.
Everything an agent needs to finish a job, and everything you need to check it.
A real editor, with an undo for everything
- See every change as a diff, per file or since the last commit.
- Checkpoints before and after each agent turn. Undo, redo, or go back to any message.
- Terminal errors get a “Fix with agent” button.
- Git and GitHub: commit, branches, push, pull, clone, publish.
- 100+ languages highlighted, search across files, command palette.
Agents test their own work before they say “done”
- A real browser: they open your project on localhost and look at the screenshot.
- They use it like a person: click, type, scroll, read the console.
- Desktop, tablet and phone, checked one after the other.
- You see what they saw: every screenshot stays in the chat.
Images, voice and video, made inside the project.
Pick a media model once. From then on, agents make what the project needs and save it where it belongs: a logo, a launch poster, a narrated demo.
Images in the shape you ask for
“A 16:9 launch poster at sunset.” Square, wide or tall, and it shows up right in the chat.
Voice-overs and dictation
Narration in the voice you choose. Or press the mic and talk instead of typing.
Video
Short clips for a launch or a demo, wide or vertical. Each one asks before it spends.
A plan they follow and a memory that lasts.
Two plain Markdown files in your project keep the crew on course. You can read them, edit them and commit them.
PLAN.md
Phases and tasks. Agents mark them doing, done or blocked, and add the follow-ups they find. A kick-off task writes the brief and researches competitors and the market first.
Phase 2 · Forecast screen 3 of 5 doneAGENTS.md
The project's long-term memory, read by every agent and updated after each working turn. What was decided and why, how to run it, and the mistakes not to repeat.
What the editor offers, in one list.
All of this is in the app today.
Agents
- Several agents per project, each with a role and model
- A lead agent that splits work across up to six helpers
- Agents you add yourself, each in its own chat tab
- Compare mode: two models do the same request, you keep the better result
- File ownership so agents cannot overwrite each other
- Message queue while an agent is busy
- Work continues with the window closed and resumes after a restart
- Approval before shell commands
Planning and memory
- Project plan in phases, with task status
- Kick-off: brief, competitor and market research, then a plan
- Long-term project memory, updated automatically
- Automatic summary when a chat gets long
- Skills: playbooks for quality, testing, design and security
- Ready-made tasks for common jobs
- Docs library for the project's Markdown and media
Build, run and check
- Live preview on localhost at desktop, tablet and phone size
- Browser testing with screenshots, clicks and console
- Integrated terminals with “Fix with agent”
- Run the open file or the whole project
- Checkpoints, undo, redo and restore to any message
- Diff viewer for every change
- Edit and resend any of your messages
Models and cost
- Claude, OpenAI, OpenRouter, Gemini, Groq, DeepSeek, Z.ai (GLM), Moonshot Kimi, Qwen, Together AI, Fireworks, Cerebras, Cloudflare Workers AI, Mistral, xAI
- Local models through Ollama and LM Studio
- Thinking level per agent, from Off to Max
- Tokens and cost per reply, per model and per project
- Model prices shown in the picker
- Automatic retry when a connection drops
Media and input
- Image generation in the shape you ask for
- Voice-overs with a choice of voices
- Video generation
- Dictation with a microphone picker and live level
- Attach images, PDFs and text; drag, drop or paste
- Agents can look at images in the project
Editor and Git
- Syntax highlighting for 100+ languages
- File tree, tabs, search across files, command palette
- Viewers for images, audio, video and PDFs
- Commit, branches, push, pull and clone
- Publish a project to GitHub
- Web search and page reading for agents
Three steps, then you are riding.
Open a folder
An existing project, a new empty one, or a repository cloned from GitHub. Several projects stay open at once.
Connect a model
Paste an API key for a provider you use, or point the app at Ollama or LM Studio and pay nothing.
Say what you want
The crew plans, builds, runs and tests. You watch the preview, read the changes, and steer.
No cloud workspace. No lock-in.
VibeCode.surf is a desktop app. It works directly in a folder on your computer and talks only to the model providers you connect.
| What | Where it lives |
|---|---|
| Your code | In your own project folder. Agents edit real files, and you can open the same folder in any other tool. |
| API keys | On your computer. Requests go straight from the app to the provider you chose. |
| Conversations and history | In the app's data folder on your computer, not in your project and not on a server. |
| Plan and memory | Plain Markdown files in the project (PLAN.md, AGENTS.md) that you can read, edit and commit. |
| Your account | An email, a name and a password on our server, plus which computers are signed in and the app version they run. That is what delivers updates and your plan. Your code, chats and keys are never sent to us. |
| Commands | Agents ask before running a shell command unless you tell them not to. |
Good to know before you try it.
Is this another VS Code fork?
No. It is a separate desktop app built from scratch around the chat and the agents. There is a capable editor inside it, but the centre of the screen is the conversation and the running result, not the file tree.
Which platforms does it run on?
macOS, Windows and Linux, installed with one command from your terminal. The Mac version has had the most testing so far; tell us what you hit on the other two.
What does it cost?
The beta is free. Pro is a one-time payment per computer, starting at $99 while the launch price lasts; see pricing. Either way you pay your model provider for what the agents use, at the provider's own prices, and the app shows you that cost per reply and per project. A local model costs nothing per request.
Why do I need an account?
You sign in once when the app first opens. The account is how the app receives updates and how your plan follows you between computers. It holds your email, name and password, and nothing from your projects. After signing in, the app keeps working offline.
Will an update touch my projects?
No. An update replaces the program only. Your projects are ordinary folders, and your chats and settings live in a separate data folder that every version reads. The app also keeps a copy of your settings and chats from before each update.
Can I use it without sending my code to anyone?
Yes, with a local model through Ollama or LM Studio. With a hosted provider, the parts of your project the agent reads are sent to that provider as part of the conversation, as with any AI coding tool.
What is not there yet?
Autocomplete and go-to-definition from language servers, inline AI edits inside the editor, a debugger, and extensions. It is a beta, and those are on the list.
What happens if an agent makes a mess?
Every agent turn is checkpointed before and after. Undo it, or restore the whole project to any earlier message, from the History panel.
Install from your terminal.
One command downloads the newest build for your computer and installs it. Mac, Windows and Linux.
curl -fsSL https://vibecode.surf/install.sh | sh
irm https://vibecode.surf/install.ps1 | iex
curl -fsSL https://vibecode.surf/install.sh | sh
Installs VibeCode.surf.app into Applications. Works on Apple silicon and Intel, macOS 12 or later.
Run it in PowerShell. Installs for your user and adds a Start menu entry. Windows 10 or later.
Installs an AppImage to ~/.local/bin and adds it to your applications menu.
Prefer to look first? Both scripts are short and readable: install.sh · install.ps1
Catch the wave early.
It is a beta: install it, build something, and tell us what broke and what you wish it did.
Free during the beta · bring your own API key or a local model