I finally hung up my corporate boots in December 2024. After decades as a software executive (years most recently spent as a CTO/SVP/COO), I was ready to leave the daily grind of project planning and review meetings behind. But like many technical leaders who retire, I found that my curiosity didn't retire with me.
By early 2025, I decided I wanted to build a mobile app. I couldn't code much anymore, and I didn't have the budget (or the patience) to hire a development team, but the timing seemed perfect. We were just seeing the rise of the AI Software Engineer. Coding agents like Copilot, Replit and Cursor were commonly being used. Kiro launched in mid 2025, Devin was very popular too, and newer tools like Antigravity were promising to do much more. The narrative everywhere was that if you could write a prompt, you could build a full application.
So, I tried the ones that I could in my IDE. I also tried harder with the chat LLMs. I spent a few weeks trying to make them work. I messed around with configurations of cloud-based chat LLMs with local LLMs, even trying to get multiple instances to "talk" to each other like a virtual team via Python scripts.
It didn't work. The problem wasn't the code itself. These agents could write brilliant functions or generate a clean UI component. The problem was sustaining coherence. As soon as the project grew beyond a handful of files, the agents lost the plot. They couldn't "see" or recall the whole architecture. They were excellent tactical workers, but they were terrible strategic architects. They had no memory of the decisions we made a day ago.
The problem wasn't the code itself. These agents could write brilliant functions or generate a clean UI component. The problem was sustaining coherence. As soon as the project grew beyond a handful of files, the agents lost the plot. They couldn't "see" or recall the whole architecture. They were excellent tactical workers, but they were terrible strategic architects. They had no memory of the decisions we made a day ago.
The "Manual" State Machine
I realized that if I wanted to build a complex app without a human team, I couldn't rely on the AI to be the Project Manager. I had to do it.
So, I stopped trying to make the agents autonomous. Instead, I developed a rigorous, almost mechanical protocol for myself.
I would write a specification for a single feature. I’d paste that into a chat. I’d take the resulting code, review it, and integrate it locally on my machine. But here was the crucial part: before I moved to the next feature, I would manually update the specification document to reflect the new state of the code. Then I would paste that updated context back into the next chat session. I provided coding standards and work rules.
I became the Context Manager. I treated the LLM as a stateless worker that needed to be reminded of the master plan every single time we spoke.
It was tedious, but it worked. Slowly, the app took shape. But quite early on, I switched tracks on what I was building. Instead of building my mobile app I focused on building a tool that automated the manual process I was using for development.
Building Klyve with Klyve
Over the last year, I built Klyve. It is a local-first, Automated Software Factory.
The irony is that Klyve was built using the exact process it now automates. I used that manual cycle - spec, code, integration, context update - to write the code for the tool that now performs that cycle automatically. It took me a year of "manual automation" to reach this Beta release. In that time, other tools have pushed the boundaries of what such frameworks can do, but the core problem remains the same: LLMs need a Senior Partner.
Klyve is that Senior Partner. It is a desktop application (for Windows and Linux) that orchestrates the entire SDLC. It holds the architecture in its "head" using a local, encrypted SQLCipher database, and dispatches tasks to the LLM, ensuring that every line of code serves the larger design.
Why I'm Releasing It (For Free)
I am releasing the binaries for free because I believe this is how senior professionals should interact with AI: not by letting it drive, but by building a better steering wheel.
It is closed source for now, mostly because I want to see if the utility of the binary stands on its own, and how to refine it further. It respects your privacy (Bring Your Own Key), keeps your IP on your hard drive, and helps you act as the architect while the machine does the grunt work. It also does a lot of the tedious documentation work.
If you are a solo developer or a manager who wants to get back to building without the overhead, give it a try. It’s the tool I built to replace (as much as that can be made possible) the team I didn't have.
You can download the Beta for Windows and Linux at klyve.online.
(Note: The screenshot above shows the Backlog View, where Klyve breaks down the high-level specs into actionable tasks before sending them to the agents.)
Top comments (0)