go touch grass

mobile IDE that lets you code, SSH into servers, orchestrate AI agents, get live previews, and deploy. It’s like Replit + Cursor on mobile so you can push to prod while touching grass.

Comment

Inspiration

both our laptops broke within a two week span and we had to continue shipping product for customers. we ended up shipping while taking walks by coding on our phone using termius + a local dev server on our phone browser. wanted to let others do the same easily, and make developers more productive from anywhere they are by coding on their phone. we believe more code will be written on phone than laptops in future.

What it does

go touch grass is a mobile IDE that lets you code, SSH into servers, orchestrate AI agents, get live product previews, and deploy right from your phone. It’s like Replit + Cursor on mobile so you can push to prod while touching grass.

How we built it

  • Agent tracing: Used Weave to monitor and debug our AI agents running against deployed GitHub repositories.
  • One-click cloud deploys: Built a wrapper that spins up instances on GCP, AWS, etc. straight from any repo, auto-applies code updates, and shows a live preview.
  • MCP orchestration layer: An MCP server talks to the Google Cloud API, parses the user’s request, launches an agent that writes the startup script, and boots the instance.
  • Live preview plumbing: Front-end iframes capture the instance’s public IP and pipe edit commands to remote coding agents, so changes appear instantly in the browser.

We have an MCP Server that does:

  • Startup script boots Debian VM
  • Apt upgrades, installs core packages
  • Docker Compose, Node, Nginx ready
  • OpenAI key exported for Codex
  • FastAPI template cloned under /opt
  • docker-compose launches backend frontend services
  • Nginx proxies / and /api
  • FastCGI endpoint /agent wired up
  • fcgiwrap executes codex CGI wrapper
  • Wrapper decodes querystring to args
  • Codex CLI processes request command
  • Codex logs metrics via WandB
  • Weights-Biases Weave stores execution traces
  • Command output streamed to client

Challenges we ran into

allowing multiple projects with different tech stacks from github to generate a standardised live preview

Accomplishments that we're proud of

The ability to kickstart remote servers using an MCP server on cloud providers like Google Cloud and having a live feedback on the custom app that is being deployed.

What we learned

We learned that sticking DevOps with development and coding agents in one app makes it easy for anybody to link any GitHub repo, kickstart their web app, having live previews, and being compliant to deploying to cloud providers as easy as possible.

What's next for go touch grass

Next thing is a broader integration with other cloud providers since right now we only use Google Cloud and see how we can orchestrate multiple AI agents around one code repo to see what are the results of it.

Built With

Share this project:

Updates