Dyad ships as a free, open-source local AI app builder for Linux, MacOS and Windows

Dyad is described on its GitHub page as a local, open-source AI app builder positioned alongside hosted services like Lovable, v0, and Bolt but running on the developer's own machine. The ZDNET walkthrough documents installing it on Pop!_OS Linux via DEB, RPM, or AppImage, with parallel installers for MacOS and Windows. After installation, the user adds an AI model through the Model Providers settings, either selecting a built-in option or wiring in a custom endpoint for local inference. The output is a web app that can be published to GitHub, cloned, and run locally, meaning the source ends up under the developer's control rather than inside a vendor tenant. The author reports finishing a functional app in two days of elapsed time, though only because OpenRouter's free quota ran out partway through and the build had to be retried the next day.

Tagging this under security is defensible because the central trade-off is where prompts, source code, and intermediate artifacts live. A cloud-hosted builder sees your specification and the generated code on its servers; Dyad keeps both on the workstation and only forwards the prompt to whichever model provider is configured. That makes it viable to plug in a locally hosted model for fully air-gapped builds, which is the configuration the author flagged as the primary motivation. Engineers evaluating the tool still need to audit what each chosen model provider retains, since switching to a free cloud endpoint like OpenRouter re-introduces the same prompt-leak surface that motivated the local-first choice in the first place.

The free build path is functional but rate-limited: the author hit OpenRouter's free quota mid-build and relied on Dyad's resume-from-checkpoint behaviour to finish the next day. For predictable turnaround, the Pro license at $20 per month adds model breadth, an advanced agent mode, and auto-debugging, modest compared with hosted AI builders but not free. Local-model execution is gated by hardware, since Dyad itself does not ship a model and the author points readers to a separate CodeGPT guide for picking a coding-capable model that fits the host. None of those trade-offs are benchmarked in the article, so they read as observed rather than measured.

Only ZDNET carried this item, so there is no corroboration on whether the two-day build time or the OpenRouter retry behaviour is typical across users or model choices. Readers should treat the reported experience as a single developer's walkthrough rather than a benchmark, and weigh the 'ZDNET's key takeaways' framing as opinion rather than independent evaluation. The piece also stops short of comparing Dyad against the cloud-hosted builders it is positioned against, so the privacy claim rests on architecture rather than any audit.