OVM - Open Version Manager

Inspiration

I found in the early days of coding agents that different versions would tweak the system prompt and often end in quite different results. Due to this, and also discovering that my terminal launches were slow due to using NVM and promptly switching to FVM and going from 3s to .2s I thought it would be cool to build an extensible version manager for coding agents.

I also build a plugin system that lets to diff the oss agents (or whatever you want to - not legal or financial advice) - which helps me see if any of my coding agent adjacent projects will be impacted by changes made in the coding agents or agent SDK's.

What it does

Manages autoupdates for Coding Agents (or keeps them stable) and allows you to promptly switch between them. It also has benchmarking capability to evaluate trends in the CLI's.

So far, it supports Claude Code, Codex, PI and ... Claudex 👀

How we built it

Lots of Claude and Codex honestly, I agree with Stienberger and live polyagentamarous - I can't really say how much each did, but they both did a lot, sometimes Codex leads and Claude Executes, sometimes Claude leads and Codex executes. Sometimes I lead and they both execute.

As with all things in the agentic engineering era, I've been adding too many features and polishing too much before go live, but I've never maintained an OSS project before properly, so I'd rather try my very best to get the security and experience totally down before pushing slop to the public.

Challenges we ran into

Honestly mainly my time and attention, and beyond that the polish that I desire for a public facing product. I've been running OVM for myself for many months now, but I've been storming it these past few weeks to get into a shape that I'm finally proud of.

Accomplishments that we're proud of

This title was conveniently timed. I'm proud of the whole thing honestly, it slots neatly into a dev tool ecosystem that I'm building around myself trying to make shipping ios/macos/ whatever else I want to less painful and more fun.

I'm especially proud of the benchmarks that come out of this that show a convergence from both Codex and Claude Code in package size and overall performance, and also the surprising discovery that although Codex init's insanely fast, time to first response is still won by Claude Code, and I think it's interesting because although I see people online talking about Codex being faster overall for tasks, I do find myself reaching more for Claude Code to get things done due to it just feeling snappier likely due to time to first response.

Claudex however ... well that might help to solve that, but I've spent more time in adding it as a product to OVM rather than testing it out as a go to harness.

Finally, I'm proud that the changes that came to Codex for MCP loading being blocking were actually something that I was running on my own fork locally for months before it was fixed upstream, and that I think is what I'm most proud of with OVM.

What we learned

A bunch about how to manage versions securely; also that in fact, a lot of existing version managers work less securely .. and finally, that I was able to put metrics behind the general vibe that had me picking claude code > codex - as this was due to time to first response which makes so much sense.

I love using codex also and swap between them frequently, and hopefully with these insights its something that can be improved :)