That friction is the centerpiece of a hands-on evaluation published on November 22, 2026, by XDA Developers writer Parth Shah, who argues that an underrated open-source desktop app called Dyad captures the original promise of vibe coding better than heavyweight coding agents like Google Antigravity, Anthropic’s Claude Code, and OpenAI’s Codex. While those flagship tools shine when orchestrating sprawling engineering workflows, Shah found himself craving a simpler desktop experience—one where plain English turns into a live, clickable preview without living inside a terminal.
Dyad is an intriguing proposition for Windows power users, indie builders, and IT tinkerers alike. Still, separating the marketing magic from the technical reality requires looking past a slick prototype—especially when it comes to Windows setup quirks, what "local-first" actually protects, and where free AI tiers quietly hit a wall.
Why Heavyweight Coding Agents Can Feel Like Overkill
To understand Dyad’s appeal, it helps to look at the crowded 2026 landscape of AI development tools. Platforms like Google Antigravity, Claude Code, and OpenAI Codex have surged in popularity by acting as agentic collaborators capable of planning multi-step edits, executing shell commands, and refactoring entire codebases.
That power comes with cognitive overhead. As Shah noted in his XDA Developers piece, firing up a full agentic environment often means babysitting terminal output, managing project scaffolding, and clicking through endless permission prompts. When you are building a production Android app or wrangling enterprise microservices, that rigor is a feature. When you just want to mock up a dashboard before lunch? It can feel like commuting to the corner grocery store in an eighteen-wheeler.
Dyad approaches the problem from the opposite angle. Built as an Electron-based desktop application for Windows, macOS, and Linux, it positions itself as a local alternative to browser-based AI app builders such as Vercel’s v0, Lovable, Bolt.new, and Replit. Instead of renting a hosted sandbox, you describe what you want in a chat pane, review the generated code, and watch the app spin up in an integrated preview right on your PC.
Inside the "PitchFlow" Experiment
To test whether Dyad could handle a realistic personal workflow, Shah used it to build PitchFlow, a productivity app for tracking article pitches, active drafts, and publishing schedules. Rather than tweaking config files, he fed Dyad a roughly 100-word natural-language prompt asking for a calendar view, colorful analytics charts, quick idea capture, smooth animations, rounded cards, and dark mode.
According to Shah’s account, Dyad responded more like a collaborative product designer than a rigid code generator—asking clarifying follow-up questions to refine the scope before rendering a preview he could iterate on through conversational prompts. He emphasized that PitchFlow made the most sense as a fast client mockup or proof-of-concept before committing to heavier tooling. Crucially, Shah’s impressions reflect one reviewer’s experience rather than an independently benchmarked test of build reliability or feature completeness.
What "Local-First" Actually Means on Windows (and What It Doesn't)
Dyad’s headline pitch—a "free, local, open-source AI app builder"—invites an easy misconception: that everything happens privately offline on your hardware. The reality is more nuanced, and Windows administrators or privacy-conscious developers should draw a bright line between local code ownership and local AI inference.
- Local code and zero lock-in: Your generated project isn't trapped inside a proprietary cloud builder. Dyad writes standard web code—typically React or Next.js with TypeScript and SQL—into a regular folder and Git repository on your drive. If you outgrow Dyad, you can close it and open that exact folder in Visual Studio Code, hand it over to Claude Code, or push it straight to GitHub.
- Cloud APIs still leave your PC: Unless you explicitly configure a local model, Dyad is only the orchestrator. When you plug in an API key for OpenAI, Anthropic, Google Gemini, OpenRouter, Azure OpenAI, AWS Bedrock, xAI, Google Vertex AI, or MiniMax, your prompts and relevant source context travel over the wire to that provider.
- Dyad Pro proxies traffic: Dyad also sells optional paid tiers—Dyad Pro at $20/month (200 AI credits) and Dyad Max at $79/month (900 AI credits)—so users don't have to manage third-party API keys. Under Dyad’s privacy policy (updated June 22, 2026), Dyad Pro proxies prompts and code through Dyad’s own servers to reach hosted models. Dyad states it never trains its own models on user data and prohibits provider training except on specific cost-effective models explicitly flagged in the UI as allowing provider training.
- Free cloud tiers carry data caveats: Dyad’s model documentation warns that Google Gemini’s free API tier permits Google to review and use submitted prompts to improve its products, and free models aggregated through OpenRouter commonly allow providers to train on your inputs.
Want true air-gapped privacy? You can pair Dyad with Ollama or LM Studio to run open-weight models locally on your own GPU or CPU. Just be realistic about Windows hardware limits: Dyad’s own documentation cautions that smaller local models can struggle to follow its coding instructions, and independent hardware testing notes that getting solid coding output locally often demands hefty memory—such as a 24 GB graphics card on a PC or 32 GB+ of unified memory on a Mac.
Setup Realities: Not Quite Zero-Click
Even though Dyad skips terminal-first workflows, Windows users shouldn't expect a magical single-click install with zero technical awareness. Dyad’s official quickstart documentation notes three practical hurdles:
- Microsoft Defender SmartScreen: Windows users launching the installer may encounter Windows' familiar blue SmartScreen prompt and need to click More info followed by Run anyway. (Because Dyad’s core codebase is open source under Apache 2.0—with
src/profeatures governed by the Functional Source License FSL-1.1-ALv2, which converts to Apache 2.0 after two years—security-minded users can inspect the repository before running binaries.) - Node.js dependency: Dyad requires Node.js installed on your Windows machine to serve and run the JavaScript/TypeScript applications it generates.
- Code approval gate: Submitting a prompt does not silently execute arbitrary files. After the AI finishes generating code, Dyad requires you to review and approve the changes before they are written locally and launched in the live preview pane.
For a safe first test on a Windows PC, start with a throwaway non-sensitive project, make a small UI change, read the generated diff before clicking approve, and keep confidential enterprise code off free-tier cloud keys.
Free vs. Paid Models, UI Friction, and Current Release Pace
In his XDA Developers assessment, Shah raised two fair criticisms: first, that Dyad’s desktop interface feels overly busy—crowded with icons and live RAM/CPU usage meters compared to leaner alternatives like OpenCode (an MIT-licensed terminal, desktop, and IDE coding agent built by Anomaly)—and second, that getting consistently strong output still pushes users toward paid models or API credits.
| Aspect | Dyad (Desktop App) | Cloud Builders (Lovable / v0 / Bolt.new) | Coding Agents (Claude Code / Codex / Antigravity) |
|---|---|---|---|
| Primary Interface | Local GUI chat + live preview | Browser-based cloud workspace | Terminal CLI / IDE extensions |
| Where Code Lives | Local folder / Git repo on your PC | Vendor cloud (with GitHub sync options) | Local workspace or cloud sandbox |
| AI Model Flexibility | 9 built-in cloud providers, custom endpoints, ChatGPT/Claude subs, or Ollama/LM Studio | Mostly bundled/managed by platform | Provider-tied (Antigravity supports multi-model selection) |
| Pricing Model | Free desktop app (BYOK/local); Pro $20/mo, Max $79/mo | Free daily/monthly caps + paid tiers | Usage-based API or developer plans |
| Best Fit | Rapid prototypes, visual mockups, local ownership | Zero-setup web deployment | Multi-file engineering, complex refactoring |
On the cost front, official Dyad documentation adds welcome nuance to Shah's critique. You can use Dyad without spending a dime by pairing it with Google Gemini’s free API tier, OpenRouter’s free models, an existing free ChatGPT account, or local models via Ollama and LM Studio. Still, Shah’s broader point holds true in daily practice: free cloud tiers impose strict rate limits and privacy trade-offs, while small local LLMs frequently trail frontier paid models when generating multi-component interfaces.
Meanwhile, the project’s release cadence shows active Windows engineering. Checked on November 22, 2026, Dyad’s latest stable GitHub release is v1.23.0 (published November 9, 2026), which resolved several nitty-gritty Windows pain points—including image uploads failing when hidden or system .gitignore files are present, and stopping Git from deleting through directory junctions when cleaning up isolated workspaces. A prerelease build, v1.24.0-beta.2, followed on November 20, 2026, introducing startup and navigation stability improvements alongside Windows-focused CI and pull-request validation fixes.
From Shiny Prototype to Production Reality
Dyad isn't limited to static front-end mockups. It integrates directly with GitHub, Supabase (supporting authentication, databases, and server functions), and Neon, and even includes an AI_RULES.md mechanism to enforce coding standards alongside an experimental AI Security Review panel.
Yet both Shah and Dyad’s own documentation wave the same yellow flag: a working live preview is not production-ready software. Dyad explicitly warns that its experimental Security Review feature may miss vulnerabilities, and its AI_RULES.md guide stresses enforcing backwards-compatible database changes before going live. Before any vibe-coded prototype touches real user data or company credentials, developers still need to audit authentication flows, row-level database permissions, error handling, and deployment hardening the old-fashioned way.
The Verdict for Windows Builders
Does Dyad dethrone Google Antigravity, Claude Code, or OpenAI Codex? Not for heavy-duty software engineering—and even XDA Developers’ Parth Shah admits he isn't deleting those tools anytime soon. What Dyad nails is the middle ground that agentic CLIs often ignore: a visual, file-owning, model-agnostic desktop studio where Windows users can spin up a Next.js prototype in minutes and walk away with clean, portable source code.
If you want to keep your code on your own SSD without paying a cloud-builder tax—and you don't mind installing Node.js and keeping a close eye on API keys and security audits—Dyad deserves a spot in your Windows dev toolkit.
Key Takeaways:
- Ideal for prototyping: Dyad lowers the barrier for visual mockups and side projects without terminal-heavy workflows, though complex production builds still require manual engineering and testing.
- Own your code, watch your prompts: Project files stay locally on your PC for easy editing in VS Code, but cloud API keys still transmit prompts externally unless you run local models via Ollama or LM Studio.
- Active Windows support: Stable release v1.23.0 and prerelease v1.24.0-beta.2 show steady Windows-specific bug fixing, even if the desktop UI could still use some decluttering.
References
- Forget Google Antigravity, Claude Code, and Codex, this underrated tool is the perfect example of vibe coding XDA · 2026-11-22T11:00:21+00:00
- github.com github.com
- Local models | Dyad dyad.sh