Skip to content
#AI

Dyad - local-first AI app builder: like Lovable/Bolt, only on your computer (AI column)

#AI #Engineering #Software #Architecture #DistributedSystems #Agents #ML

Dyad Local‐first AI app builder: like Lovable/Bolt, only on your computer (Rubric #AI)

I told you about the product once. Lovable, a kind of "AI app builder", in which he wrote a request → received code changes → saw a preview → published for sale. In fact, Dyad does exactly this cycle, but locally. In fact. repositories Dyad is a local/open-source alternative to Lovable/v0/Bolt. (This is an alternative approach to cloud platforms.).

I’m also talking about the local-first approach. toldWhen doing sammari on the documentary and in dyad there are all signs of this approach.

  • Your code and project are on disk.: without "platform cage", you can import/export and freely switch between Dyad and conventional dev-workflow
  • Bring your own keysYou connect the right providers and models (OpenAI/Google/Anthropic etc.), without forced vendor-lock-in
  • ExtensibilityYou can add tools through MCP servers and live on patterns (templates) stacked

Dyad functionality is similar to Lovable and includes

  • Full-stack applications via SupabaseQuickly connect database, auth and server functions
  • Security review based on AI: run a check, get finds with criticality levels and can ask AI to fix a specific problem; check rules can be tweaked through SECURITY_RULES.md
  • Automatic versioning: Each AI edit is a new version; it is essentially a git commit. You can open a list of versions and “restore” to the last successful state. (mechanical git revertBut history is not lost.)
  • Publication"Publish anywhere" - GitHub/Vercel or your stack

When it comes to architecture, it is as follows: Dyad is electron-application:

  • UI lives in renderer process, Access to the file system and "application of changes" main process,
  • Communication goes through IPC.

The key trick of AI-editing is that the model responds not with “just text”, but with a XML-like command format. UI streams and beautifully displays the answer and then the main process. (response processor) applies commands to the project: create / modify / delete files, add npm packets, etc. Separately, it is noted that Dyad does not strive to be as “agent” as Cursor, because complex agentic cycles quickly become expensive – so more often it is “one request → one neat pass”, with an optional auto-fix TypeScript errors.

You can read more about the architecture of the project on the website system-design.space, where I posted detailed analysis of pictures and visualizations.

#AI #Engineering #Software #Architecture #DistributedSystems #Agents #ML