Typemill 2.23.0: Bring Your Own AI

Typemill 2.23.0 is a small release with one major change: the AI integration with Kixote has been completely refactored. You can now connect not only Claude and ChatGPT, but also any service with an OpenAI-compatible API. That means local or self-hosted setups like Ollama, LM Studio, or Opencode should work as well.

Full Flexibility With AI

The new AI integration is a bit more technical, but it makes the system much easier to maintain. For example, you no longer have to wait for a new Typemill release to use a newly released model. You can simply change the name of the default model in your settings. There is also a new info tab, so you can quickly switch between different models while editing content in the Kixote interface.

Please note that this is a BREAKING CHANGE and requires a new AI setup. The old settings in the AI tab will no longer work after switching to v2.23.0.

The Trend to Local and Privacy-Focused AI

The request for Ollama support came up frequently, especially from business users who run Typemill internally and want to keep their data and workflows private.

I also recently moved from ChatGPT and Claude Code to Opencode. Opencode is a Claude Code alternative that not only offers a wide selection of models, but also lets you connect to locally hosted models with Ollama or LM Studio. So the move to a generic connector felt like a natural next step. I also think that shifting to AI in general, and to local models in the long run, opens up great opportunities to run small open-source projects more sustainably in the future...

Financial Aspects of Typemill and AI

Typemill started as a small side project in 2017 and grew into something closer to part-time work. But a project like this is very time-consuming and expensive. Since the move to commercial plugins in 2024, income has increased slowly, but it is still not enough to fund the time I would like to spend on it. My goal is to increase income from the current 250 MRR to about 800 MRR in 2027.

AI can help here. Assisted coding can significantly increase output and quality if used well. But tools like Claude Code can easily eat up 100 to 200 euros each month when used for heavy and complex coding tasks. With OpenAI-compatible services, locally hosted models, and more thoughtful workflows, this becomes more realistic if the transition is handled carefully.

Next Steps

In the next releases, I want to focus on two main tasks, alongside the ongoing integration with LDAP, SAML, and OpenID: adding more useful AI features to Typemill and setting up Git integration.

Git has been requested many times and has an interesting use case. With AI, many people and companies realize that documentation inside repositories is often the easiest way to manage software documentation. This is usually called docs as code. However, this documentation is often hard to access for editorial teams and spread across multiple repositories. Typemill could become a natural layer between repositories and editorial teams.

There is another perspective: Markdown is used not only for documentation, but also for user stories, prompts, plans, agents, skills, and test scenarios across the software pipeline. You could call it code as content. And this is where a Markdown-based CMS could play a surprisingly useful role in the future...