DEV Community: Manish

Small win: Made CodeRunner and got appreciated by Google Gemini team!

Manish — Wed, 02 Jul 2025 05:29:06 +0000

A small story to motivate you guys to work on cutting edge frontier tech:

You might have heard about Gemini CLI[1] release a few days back. I managed to integrate Apple Containers (also released recently) to execute the code generated by AI safely.

Gemini CLI supports code execution itself but via shell, which can be a security issue. CodeRunner solves this by running it in a safe isolated container.

Got appreciation from Google team:

cperry is from the Google team who commented on my reply with "<3 amazing". I will consider it a win.

Some asked me,

That is amazing. Could you tell me how you did it and how you even thought of it in the first place? I'm a student so please eli5

Context: Everyday something new is released, I read about it and in depth -- doesn't take a lot of time now as I have been doing this for a few years.

When Apple Container was released I had already read about WASM[1], docker[2] etc. The native apple containers are kind of a big thing I felt, so I and my friend created CodeRunner[4] for apple containers 4 days back.

I had also seen a trend of MCPs[3] -- initially I brushed it off but it seems there are a lot of tech built around it. So, converted my APIs (which eventually execute code) to MCP APIs which do the same thing but are now MCP compatible.

Cut to Gemini CLI release, I went to check if they have MCP support and since I had read about MCPs before I knew it would be a quick change to make it work for Gemini and indeed they did have support and after that it was a matter of config change!

Moral of the story is, be ready, opportunities will come and then you strike.

And if somehow we magically reach 500 stars on github -- who knows it might become part of official Gemini CLI. Lot of small wins can lead to a big win!

WASM: https://en.wikipedia.org/wiki/WebAssembly

docker: https://en.wikipedia.org/wiki/Docker_(software)

MCP: https://en.wikipedia.org/wiki/Model_Context_Protocol

CodeRunner (mine): https://github.com/BandarLabs/coderunner

Gemini is so good, I have let it control/use my phone

Manish — Mon, 20 Jan 2025 14:35:56 +0000

Click3: LLMs have taken control of my phone

Draft a mail to someone@gmail.com and ask for lunch next Saturday.

Find out what's my rating in Uber?

Start a 3+2 game on lichess

All of these tasks require you to know the UI, like where to find the ratings - "is it in the profile section or is it even there in the app?" and do multiple clicks afterwards.

But with this framework, you can just type in plaintext and see the LLM do it for you.

How does it work?

It has three separate components, as seen below and each have their own separation of concerns:

Planner: Plans the next step (given current screenshot and previous actions)
Finder: Finds specific UI elements (whatever Planner asks it to find)
Executor: Clicks, scrolls, types etc

OpenAI / Gemini / Local LLM

I have added supports for all of the above, to be precise, the following is the recommended models for each component.

You can use your own keys and run it locally, the support for local LLM is what I am most excited about and the Molmo MLX for MacOS is a great start, I feel.

Open-sourced

This project is open sourced for everyone to use and contribute. You can also check out more demos in the README:

https://github.com/BandarLabs/clickclickclick

What do you guys think?

What are the use cases you can think of for this? For starters, I think it can be used to "create overlays of walkthrough over any app" or "automate testing of any functionality of an app" for developers.

🎙️I built "GitPodcast" to help devs understand 💡any GitHub repo easily

Manish — Fri, 17 Jan 2025 07:04:12 +0000

Understanding any new codebase was a daunting task for me while listening to any podcast is something that comes very naturally to me. So, what if we combine the two ideas - that is understand any GitHub repository via a Podcast.

GitPodcast

Gitpodcast: In this tool you can enter any GitHub url, and get the podcast within a minute.

Users Feedback

Users wanted longer podcast, so I have added the ability to get upto 10 minutes of podcast (Do people want even longer podcast or is this enough?)

Talking to users really helps you understand what exactly they want, and this was the main feature they wanted instead of the options to add multiple voices which I was thinking as the main missing feature.

Got the top 3 product award from producthunt on the day of the launch.

How to use

Two ways to use it:

Just replace hub in the url with podcast (when you are at any github repository)
or Paste the repo link in the website input box

What do you guys think?

You can check it out here: https://gitpodcast.com

Open-source

The code is open source and you can host it yourself with your keys - https://github.com/BandarLabs/gitpodcast

(In case you are logged in to github already, please do leave a star there!)