Dyad is the local answer to Lovable / v0 / Bolt: describe an app, get runnable code, keep the files on your disk. Open source (dyad-sh/dyad, ~22k stars), desktop-first for macOS and Windows, and aggressively BYO-model — OpenAI, Anthropic, Google, plus Ollama and LM Studio when you want inference to never leave the machine.
Own the code, pick the brain
Most “AI app builders” rent you a sandbox and a meter. Dyad’s pitch is the opposite: real exportable projects (think standard Next.js-style output), import existing repos, edit in VS Code when the AI stalls, deploy to GitHub/Vercel or your own cloud. No project-count tax for staying local. MCP hooks for browser/devtools/custom servers if you want the builder to reach beyond a single chat box.
That model flexibility is why it lands next to Cursor and Codex in my stack notes. I don’t need another walled garden with one vendor’s weights. I need a UI that can point at Claude for a hard UI pass, then flip to a local Ollama model for private iteration on client-shaped prototypes — without re-uploading the whole idea to a SaaS builder every time.
What “local” buys you
Files stay on disk. Offline-capable when the model is local. Built-in security scan before you ship (injection, auth misconfig, the usual). One-click Supabase/Neon for schema + auth when you’re past the static mock. Recent releases have been grinding on agent-mode context retention — the boring reliability work that matters more than launch demos.
Cyber Sentinel covered ZDNET’s Jack Wallen build: a real sibling-facing web app in about two days, mostly limited by OpenRouter free-tier quotas rather than Dyad itself — Retry the next day, finish, publish to GitHub. That’s the profile that makes this interesting beyond hobby demos: people who refuse SaaS lock-in and still want a no-code-ish path to a working front end.
Where it fits for network/security builders
I’m not replacing Cursor Projects or Codex with Dyad. Different job. Dyad is for spinning product-shaped UIs and full-stack starters without pasting yourself into a hosted builder that trains on your prompt stream. Pair it with Ollama/LM Studio when the prototype shouldn’t leave the LAN; bring frontier keys when you need better taste on layout and copy. Export the code, open it in Cursor, keep shipping.
If you’ve bounced off Bolt/Lovable pricing or privacy, try the free local path first. Star the repo, install the desktop app, point it at a model you already trust, and see whether the exported code is something you’d actually maintain. For EtherNinja readers living in agents + local AI: this is the open-source “app factory” slot that finally matches the rest of the stack — local-first, key-agnostic, and honest about ownership.
Sources: Dyad · GitHub · Cyber Sentinel coverage