Helium Agent

This is a submission for the GitHub Finish-Up-A-Thon Challenge

What I Built

What is Helium

Helium is a terminal based local AI agent that does everything a propriety agent does. It can code, research, execute bash, manipulate files, and more. It is built with a lightweight harness so that only things that you need are eating your ram.

How it started

I was absolutely stunned by how good claude code and other big tech agents were. It felt almost magical but then when openclaw came into the light I got to know that under the hood it was run by Pi. I looked around about it and found why it was being used. I was inspired by a speech of its creator telling to embrace the simplicity.

I thought to myself why not built something that lives between the simplicity of Pi and throughput of claude code. Hence, Helium was conceived.

I started coding it in python, keeping the stack as minimal as possible. I also added a voice wakeup command but scratched that because I couldn't get the voice recognition work properly without using a large model contradicting the entire lightweight claim.

This was when I took a break from it.

Helium is the biggest project I have ever made and worked on. I spent days looking for free LLM providers because my mac was not good enough for medium models and the small models were not satisfactory. Yes, I tried quantization but it was meh.
I was lucky to get one and from there it was no looking back.

Demo

The github repo is here: https://github.com/DebmalyaSen34/helium-agent

I made a video walkthrough here:
Watch Demo video

Screenshots:

The Comeback Story

Helium was just a few python files with no UI, no orchestrator, but only few tools that couldn't talk to each other. I lost the motivation to continue it and went to my home town. That time away from it made me complete it even more than when it started.

I have had a lot of free time on my hand recently so I thought why not continue Helium. I was also motivated the AI Engineer event in Europe to test new things. I continued researching on the harness development, context management, memory, tool execution and web search.

Features added

The first feature that I added was a web_search because I always fancied a agent that could utilize the internet. So, I used scraping libraries and duckduckgo search api. I tested it heavily like asking about very recent news in sports and geopolitical areas. Soon many tools and features followed:

• Research/Deep research
• Bash execution
• File manipulation
• Coding workflows

Moving on from past

I had to remove the entire voice suite because it was becoming heavy due to models being used for voice recognition and complex to maintain.

I also removed the web ui because it was difficult already to maintain the terminal operations so I scrapped that.

Removing these features allowed me to completely focus on one task. Giving a beautiful experience in terminal.

My Experience with GitHub Copilot

We live in an era where you are limited only by your imagination and not the technology.

I have a student plan for Github Copilot. Every idea was passed through it and double checked to make it is serving its purpose. Copilot wrote all the extensive tests to make sure the agent behaves as intended.

Github Copilot contributed heavily to the memory management and coding workflow implementation as I was unsure which directions to take. It and I conversed for long discussing different architecture designs and chose the one best suited for my constrains.

It would create detailed architecture-design.md files which I would review and then move on to the implementation.md file where all the magic happened. These files were in detailed and contained all the edge cases for testing. I didn't expect this level of consistency and quality from any agent.

Bottom line: without Github Copilot it would've taken few months to make it work and few more to make it not break.