K2Coder is a Mac app that builds software from a description. You tell it what you want — a website, a React web app, a Node API, an iPhone app — and it creates the project, runs it, and shows you the result inside the app. It is free to download from quantumaccordbg.com/k2coder, signed and notarised by Apple, and runs on macOS 14 or later on both Apple silicon and Intel Macs.
This post is about what it does, what it deliberately does not do, and how it treats the AI keys you give it.
What do you actually get?
Projects, like Xcode. New project, pick a kind — website, React web app, Node API, or iPhone app — and describe it. K2Coder writes the files and runs the result with one button. For React web apps and Node APIs, the AI installs packages with npm, so npm must already be installed on your Mac; K2Coder includes Node.js but not npm. A Preview pane shows it live.
It checks its own work. This is the part that separates it from pasting code out of a chat window. After building a website, K2Coder opens it in a real browser at phone width (390 px) and desktop width (1280 px), measures problems such as text or tap targets that are too small and content wider than the screen, and fixes what is really wrong; a React web app gets the same review while it is running. The check needs Google Chrome, Microsoft Edge, Brave or Chromium on your Mac, and runs when K2Coder's own engine does the work, not when Claude runs through your Claude plan or Gemini through the Gemini CLI. The phone-width check measures content that is wider than the screen, text that is too small and tap targets that are too small, so those layout problems can be caught on your Mac before a visitor sees them.
iPhone apps. SwiftUI projects that build with Xcode and run in the iPhone simulator, with the live simulator screen inside K2Coder. You need Xcode installed for this; the app tells you if it is missing rather than failing quietly.
It remembers your project. Each project keeps a small wiki of the decisions and conventions made along the way, so the next session on K2Coder's own engine starts with the project's notes, including where things live, instead of re-deriving your folder structure.
Many AI models, your choice. Gemini, OpenAI, DeepSeek and Kimi, each with your own API key from that provider; local models running on your own Mac through Ollama, which you install separately; and Claude through your own Claude plan, using Claude Code installed and signed in on your Mac (it is not included with K2Coder). You are not locked to one vendor, and you can move a project between them.
What does it not do?
It does not host anything. K2Coder builds and runs on your Mac; putting a site on the internet is a separate step, and the app does not pretend otherwise.
It is not finished for Windows. A Windows build is being made now. The download page takes an email address and tells you when it exists — it does not offer a file that is not there.
K2Coder does not upload your projects. They live on your disk, and K2Coder's own update checks and the crash reports you agree to send carry no project files. The code the AI works with goes to the AI model you choose. With your own keys or Claude plan, it goes straight from your Mac to that provider. With a local model through Ollama, it stays on your Mac. If you pay with Quantum Cloud credit, it passes through Quantum Accord's service on the way to the provider, and that service does not store it. Why we built it that way: why K2Coder keeps your projects on your Mac.
Where do your keys go?
Provider keys you enter in K2Coder's Settings window (K2Coder ▸ Settings…) for Gemini, OpenAI, Anthropic, DeepSeek and Kimi live in your Mac's Keychain. In the current version, a key added on the Providers card of K2Coder's Models page is saved in a plain text file in your home folder (~/.agents/secrets.env) instead; moving those keys into the Keychain is planned for the next update.
On K2Coder's own engine, provider keys are removed from the environment of every command the AI runs, and a test checks this. The exception is the optional Gemini CLI route, which gives the Gemini CLI your raw Gemini key; the app warns about this. That distinction matters: an AI agent that can run shell commands is one prompt-injection away from cat ~/.zshrc, and because no model-level defence against prompt injection is fool-proof, the safer design is for the secret not to be in any place the agent can read. K2Coder does not fully meet that yet. In Full auto mode the agent's commands are not sandboxed, so they can read files in your home folder, including ~/.agents/secrets.env and K2Coder's own gateway and account token files. What the next update changes about a project's own commands: opening someone else's project should not run its code.
Download
K2Coder for macOS, free, no account needed to download: quantumaccordbg.com/k2coder. The download page shows the current version and its size. The file is served from private storage through a link that expires after a few minutes, so a link pasted into a forum stops working, and once a newer build is published, the download button hands out only the new one. The AI K2Coder builds with is set up separately: your own API key from a provider, Claude Code signed in to your own Claude plan, or a local model through Ollama, which you install separately.
Windows: leave your address on the same page.
Who is it for?
Freelancers and small studios who need a working site or app faster than they can write it, and who would rather check a running result than read generated code. If you already know how to deploy a static website, a React (Vite) app or a Node API, K2Coder gets you to the deploy step; if you do not, the Quantum Accord website builder can build and host a site for you instead. Publishing a site there needs an account and a paid site.