Field notes / Open source

AI took away the hard part of open source

Open source always gave you most of a paid product for free, as long as you could run it. Now our AI agent runs it.

By , President & Co-Founder

Published

7 min read

James Galang's photo cut out like a paper print on a navy and violet paper collage, holding scissors that have just cut a long paper receipt with price tags hanging from it, beside a stack of cut-paper servers that a small paper robot is tightening with a wrench, with open padlocks, gears, code windows and a checklist around them

Long before the generative AI wave, I was a heavy open source user and a big part of the open source community. I thought open source tools were amazing. A lot of them had maybe 90% of what a paid solution had, and some were even better than the paid version.

On top of that, you could make it yours. You could fork it, customize it and add features that aren't usually there, without starting from scratch, because someone was already maintaining the repo. If a project is active, and especially if it's a popular one, you're building on years of other people's work for free.

The catch was always the upkeep

The biggest downside to open source was that you had to be pretty technical. You also had to do a lot of configuring and tinkering.

Most of these projects are run by individuals, not companies with support teams. They weren't always the best documented. They weren't always the most optimized. And when you needed a fix, it didn't always come fast. So for most businesses, the math came out in favor of paying for software, even when the open source version was just as good.

Then AI learned to work inside a codebase

Once AI got powerful enough to work directly with repositories and codebases, open source went from what I already thought of as a superpower to something even bigger. The setup, the configuring, the reading through thin documentation, the waiting on a fix: an AI agent can do all of that now.

AI takes away a lot of the downsides of open source and keeps all of the upside. You still get the free software, the control and the ability to change it. You just don't need someone on staff whose whole job is keeping it running.

Here's what that looks like at Galang AI today.

What we needWhat we used to pay for or rely onThe open source tool we run now
CRMHubSpotTwenty
Code hostingGitHubForgejo
Tunnels to our appsCloudflare TunnelsPangolin
Operating systemWindows or macOSOmarchy (Linux)
The AI agent that maintains all of itA person on staffHermes Agent

We replaced HubSpot with Twenty, an open source CRM

Here's a real example from Galang AI. Instead of paying for HubSpot, we run Twenty, an open source CRM. Its makers call it "the open alternative to Salesforce, designed for AI."[1]

If you've used HubSpot, you know how it goes. The costs climb fast, and they always get you because the next feature you want is in the next tier up. Before you know it you're paying thousands a month, and bigger teams can hit tens of thousands. There are setup fees too. HubSpot's own pricing page lists a required one-time onboarding fee of $3,000 for Marketing Hub Professional and $7,000 for Enterprise, on top of an Enterprise plan that starts at $3,600 a month.[2]

We built our workflows around Twenty, and we haven't had to touch it since. Our agent manages that tech stack, our infrastructure and the servers it all runs on.

The AI that can hack can also secure

You hear a lot in the news right now about how good AI models, Anthropic's Claude models in particular, have gotten at cybersecurity and hacking. Well, it works the same on the other side. A model that knows how to break into a system also knows how to secure one.

Anthropic is betting on that side too. When it brought Claude Mythos 5, its most capable model, to more defenders in August 2026, it also launched a $35 million fund in Claude credits for organizations working to patch vulnerabilities in open source projects.[3]

That's what AI has been doing for us. Our agent does maintenance on our open source stack every week, and security patches go on every day.

One caution. You can't go all in on open source without proper security procedures, governance and compliance, or you'll just be vulnerable. But it isn't hard to learn, and you don't have to do it alone. If everyone is saying Claude is so good at hacking, it's also smart enough to know what to secure, and so are the other frontier models.

What else we run ourselves

The CRM isn't the only piece.

Our own Git server instead of GitHub

We host our code repositories on Forgejo, an open source Git server, so we aren't relying on GitHub. GitHub has gone down a few times. In August 2026 it had an outage that lasted 7 hours and 47 minutes. GitHub said monthly commits had grown from 1.4 billion to 2.9 billion since April, and "that growth explains the pressure on our systems."[4] A lot of that new code starts with AI. By September 2026, more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March.[5] There has also been talk of GitHub charging more. In December 2025 it announced a new per-minute charge for GitHub Actions run on your own machines, then postponed it after developers pushed back.[6]

Pangolin instead of Cloudflare Tunnels

Tunnels let you put an app on the internet without opening up your server. We used to use Cloudflare's, and we switched to Pangolin, an open source version, after we ran into problems with video and large files. On its Free, Pro and Business plans, Cloudflare's terms require a separate paid service to serve video and other large files through a public tunnel.[7] We've found Pangolin to be much better.

Omarchy on our computers

We've even gone as far as an open source operating system: Omarchy, a Linux setup made by DHH, the creator of Ruby on Rails. GitHub's 2026 Octoverse report ranked it among the fastest-growing projects on GitHub by contributors.[5]

The agent that runs it all

Even the agent doing this work is open source. It's built on Hermes Agent from Nous Research. In the same Octoverse report, Hermes Agent and OpenClaw, the two agent platforms we deploy for clients, each topped GitHub's list of the most-engaged open source projects in their category.[5]

Photos and files

You don't have to start with a CRM or an operating system, either. Two of the most popular things people self-host are photos and files. Immich is an open source alternative to Google Photos, and it's really good. Plenty of people run their own cloud storage too. Either way, you're the one in control of the privacy and security.

Why would we start from scratch?

We try to be smart about how we use things. If something already exists and there's a good version of it, why would we start from scratch? Even with AI tools that can write code for us, why would we?

If someone is already maintaining it, managing it and spending their own tokens to build something amazing, why not use it and start from there? Especially when we don't have to pay for it. And if there's ever something we really want that it doesn't have, we just build it in ourselves.

There's also the transparency. With open source you can see exactly what's in the codebase and what's going on. Nothing is hidden: if a tool collects data or serves ads, it's right there in the code, and you can turn it off. You have complete control over that.

How we stay AI native affordably

Open source has increased the level of productivity and the resources we have available to us. Not many people are talking about this yet, but I really think open source is the way. It's how we stay AI native, affordably, at a time when everything comes with a subscription.

It's also work we do for clients: self-hosted open source software, set up and kept running, private AI deployments, and open source agents set up and secured for a business. If you want to know which of your subscriptions an open source tool could replace, and what it would take to run it safely, book a free 15-minute intro with us.

Galang AI newsletter

Practical AI, from the field

New Down To A Science episodes, Galang AI news and our latest guides, in one short email on Mondays when there's something new. The first Monday of each month adds a longer note on what we built, what worked and what didn't.

Sources

  1. [1]twentyhq/twenty: The open alternative to Salesforce, designed for AI (Twenty (GitHub)), retrieved . https://github.com/twentyhq/twenty
  2. [2]Marketing Software Pricing (HubSpot), retrieved . https://www.hubspot.com/pricing/marketing
  3. [3]Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders (August 21, 2026) (Anthropic), retrieved . https://claude.com/blog/bringing-claude-mythos-5-to-more-defenders
  4. [4]The August 17 outage, and the work ahead (August 20, 2026) (GitHub), retrieved . https://github.blog/news-insights/company-news/the-august-17-outage-and-the-work-ahead
  5. [5]Octoverse: Commits, pushes reach new highs in 2026 as developers build with agents (October 28, 2026) (GitHub), retrieved . https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents
  6. [6]Update to GitHub Actions pricing (December 16, 2025) (GitHub Changelog), retrieved . https://github.blog/changelog/2025-12-16-coming-soon-simpler-pricing-and-a-better-experience-for-github-actions
  7. [7]Cloudflare Tunnel FAQ (Cloudflare), retrieved . https://developers.cloudflare.com/cloudflare-one/faq/cloudflare-tunnels-faq