All writing
AIAgentsCode ReviewTools

Software Got Cheap, So I Build My Own

Agents made software cheap enough to build for an audience of one. I got tired of paying for code review bots on top of the AI subscriptions I already had, so I built Gecco.

Max KellyNovember 1, 20264 min read
Software Got Cheap, So I Build My Own

Personal software is one of the big conversations this year. Agents made building cheap enough that people are making small apps just for themselves.

Most of that conversation is about apps for your life, like a custom dashboard for your sleep data.

I have made a few of those too. Heartbeat is an app I built that plugs into my Whoop. It shows me my Whoop data, and which app I'm in when I'm most stressed.

Mine is mostly tools for how I work. That is a natural place to start, and for my tools the builder and the user are the same person.

So I build my own.

I was paying for review twice

I was paying for two AI code review bots, Greptile and cubic. Both are good. Both caught real bugs in my code.

I was also paying for AI subscriptions that run frontier models. Every review those bots ran was inference I was already paying for somewhere else.

Agents had made software cheap enough that the fix was obvious. I built my own reviewer.

What Gecco does

Gecco reviews my GitHub pull requests on the AI subscription I already pay for.

Here it is:

gecco.sh

It connects to GitHub through an app, and reviews a pull request when it opens and again on new commits. It runs on a ChatGPT subscription or an OpenAI-compatible endpoint you run, and you can host it yourself. Every finding is checked against the source before it reaches you.

Each review leads with a merge confidence from 1 to 5 and a one-line reason. Then come the findings, each with its severity and its location.

That score turned out to be the most useful part, because my agents read it.

How I use it

Gecco runs on my own repos. In the Gecco repo it is part of the merge rule: a pull request does not merge until CI is green, Gecco scores it 5 out of 5, and every finding is resolved.

The agents that write the code read the findings and fix them before I look at anything.

Gecco also reviews its own pull requests. That keeps me honest about it.

Dependency updates go through the same gate. When Dependabot opens a batch of version bumps, the agents check what breaks, fix it, and merge once Gecco signs off.

Being fair to Greptile and cubic

Greptile and cubic are real products with real teams behind them.

cubic says it is the top AI code reviewer on Martian's Code Review Bench.

On my own pull requests, Gecco found more real bugs than either of them, partly because it reviews every push. It is also noisier. cubic is quieter, and plenty of teams will want exactly that.

If you run a team and want review handled for you, buy one of them. Gecco exists because I am the user, and I already pay for the models.

What changes when software is cheap

You build for one user. Every feature has to earn its place by being something you need.

Maintenance becomes the real cost. The first version is the cheap part now. Keeping it running is where the time goes, so I hand that to agents too.

You have to measure it. Personal software can quietly be worse than the thing it replaced. Running Gecco side by side with the bots on my own pull requests is how I found out where it trailed cubic.

You can throw it away. If a product ships what I built and does it better, I will switch.

Build or buy, now

My rule is simple.

Buy when the software is a shared surface for a team, and someone else's roadmap is a feature.

Build when you are the only user, and the product is charging you for something you already have.

That second case used to be rare, because software was expensive to build. It isn't anymore.

A lot of the software I use most now has exactly one user.