How we connected Vercel, our VPS, local AI, and a step-by-step GitHub guide — and what our real-world tests taught us.
Building an open-source platform is not just about writing code. It's about making technology accessible to people who may have never used GitHub, configured a development environment, or contributed to a repository before.
At MyZubster, we're working toward a decentralized ecosystem where people can share skills, collaborate on projects, and build independent services.
Today, we made measurable progress on one important part of that vision: Zorgax, our AI assistant.
Our application runs through Vercel, while part of our backend infrastructure and local AI services operate on a VPS.
Previously, requests to the Zorgax assistant could fail with HTTP 502 errors.
We worked on a signed HTTPS gateway connecting the Vercel deployment to the VPS backend.
The result:
This is an important infrastructure milestone, but it does not mean every feature is finished.
We want Zorgax to do more than answer general questions.
One of its first practical responsibilities is to help people join MyZubster through GitHub.
A new contributor should be able to say:
Hi Zorgax, help me create a GitHub profile for MyZubster.
And receive a clear first step, rather than a complicated technical explanation.
We introduced a local GitHub onboarding guide that starts with:
Step 1 — Create your GitHub account.
Visit https://github.com/signup, complete the registration, and return when ready.
The assistant also reminds users not to share passwords or verification codes.
Our first live tests revealed a problem.
Zorgax correctly started the registration process, but after a contributor supplied a GitHub username, it sometimes restarted Step 1.
We traced the issue to the onboarding rules.
We then introduced username recognition and used conversation history to identify whether the contributor had already started the registration process.
The updated assistant can now respond:
Step 2 — You've provided your GitHub username. Open PowerShell and check your Git and GitHub CLI installations.
This change was tested before deployment.
Regression test results: 4/4 passed.
The tests covered:
After applying the change, restarting the VPS service, and testing the public chat, we confirmed that Zorgax progressed to Step 2 instead of restarting registration.
Our VPS runs local AI services, including Ollama.
During the onboarding tests, we observed two response paths:
local-rule · github-guide — deterministic instructions for GitHub onboarding.local-ollama · qwen2.5:3b — responses generated by the local language model.
The test responses reported privacy_local_only, indicating local processing for those requests.
This architecture allows us to combine predictable instructions for critical workflows with a language model for more flexible assistance.
However, the distinction matters: a model-generated answer is not the same as a verified workflow step.
Our next test simulated a Windows user who had Git and GitHub CLI installed.
Instead of progressing to GitHub authentication, Zorgax switched to Ollama and began discussing the creation of a MyZubster character.
That wasn't the expected response.
The correct next step should be to guide the contributor through:
gh auth login
Then verify authentication using:
gh auth status
Without requesting passwords, tokens, or private credentials.
This exposed the next engineering task: extending the deterministic onboarding guide so that the AI doesn't lose track of the contributor's actual objective.
We are not presenting this part as complete.
Our next development priorities are:
Longer term, we want Zorgax to support the MyZubster marketplace, contributor collaboration, and independent projects — without making the platform unnecessarily complicated.
MyZubster is exploring a model where contributors can participate not only by writing code, but also by offering services, documenting knowledge, testing software, and connecting independent projects.
A useful AI assistant should lower the barrier to participation.
It should explain one step at a time, remember progress, avoid claiming that something has been verified when it hasn't, and help people become more independent.
Today we didn't finish the entire onboarding journey.
But we verified that our Production assistant can start it, recognize a GitHub username, and continue to the next step.
Small, tested improvements are how an open-source ecosystem becomes usable.
Website: https://www.myzubster.com
Zorgax assistant: https://www.myzubster.com/zorgax
GitHub: https://github.com/MyZubster-Ecosystem
We're interested in contributors who want to help simplify onboarding, improve local AI workflows, test the marketplace, or explore decentralized collaboration.
What would you expect an open-source AI assistant to do when helping a complete beginner make their first GitHub contribution?