I was at work and I forgot to hit save before 12pm. Heres' my stuff (even if it doesn't count)
Video: https://youtu.be/mEwxmvL2HUM
Text:
Inspiration
Jac builder is pretty neat, but doesn't support mobile apps. Most vibecoded apps that make thousands of MRR are usually mobile (and specifically iOS), so it makes sense to have create an agent that works for any target that Jac supports
Jacsmith solves this itch for jac, being the only entrant in the $5 billion market that truly owns its entire stack (AND compiles desktop apps too, I don't know any other AI builders that make desktop apps).
Essentially ai builder for every kind of app.
Ok so you just made jac builder for mobile. What's novel about this?
Local first architecture isn't common with most AI app builders, with good reason, but I chose this since it makes adding native functionality (like camera access / sensor access) super easy to test.
Jacsmith internally creates a App specification graph before building that gets updated throughout the development process. This way of representing the app spec means you always have
Given that Jacsmith is a desktop app, the agent spins up cloudflare tunnels once the backend is generated so that you can access the backend even from your phone (which isn't going to be on the same local network).
How we built it
Completely Jac (minus ts imports for code mirror components), used the MCP + skill + my harness
Challenges we ran into
Tauri's devex is truly horrific. I had to reload my application thousands of times over the last few days due to bugs with tauri's file watching. SSE from the sidecar was also not easy to communicate with, and it wasn't clear how I'd do it without freezing the app while response streams.
Low context models: Featherless has 32K models available, which are not great for coding, but I was still able to use it in parts of the development pipeline by scoping out the planning process. Finding the right model for the right task was a little challenging as well, since I was looking to spend as little as possible, but GPT5.3 turned out to be really great at UI generation while remaining reasonably priced.
Accomplishments that we're proud of
Creating something that actually works, and 5 shots Linkedin completely in Jac + Mobile view.
A novel architecture for AI generated application development for real apps and not just toy projects., which lets you optimize for output token usage through abstracting away commonly used functionality
Byllm powered coding agent, with each agent stored as nodes on a graph
What we learned
Learnt a ton about system design, since I had to really plan out the architecture before jumping into coding
What's next for Jacsmith
A web based view that lets you sync you local jacsmith changes + web containers so you have the luxury of making jac apps even on mobile
I'm going to continue working on this to improve the capabilities exposed by the app spec (currently just auth and payments) so a lot more of the application generation is automated. One direction I'm leaning towards is Composio integration so that I have thousands of plugins for essentially free.
Log in or sign up for Devpost to join the conversation.