Step 1
The agent proposes. You decide.
Most AI coding tools are moving toward letting the agent decide everything. Plinth goes the other way. It's an environment for experienced engineers who want to see every decision the AI makes, stop it at any point, and change it.
The agent decides
Where most tools are heading
Where Plinth stands
The engineer decides
Step 2
Every run, drawn as decisions.
In Plinth, the agent's work isn't a scrolling chat. It's a map of decisions on a canvas: what it chose, why it chose it, and what it would have done otherwise. You can stop at any node, look at what was decided, and change it. Plinth keeps every version, so changing your mind at step four doesn't cost you steps one to three.
A run, with a change of mind
Step 5: Write the change
Step 3
When you don't know the answer yet.
Experienced engineers don't know everything either. When Plinth stops at a decision you're unsure about, it doesn't just wait. It lays out what to keep in mind, what to look at, and what to research, with every input you need in one place. You can open a side conversation about that one decision, look things up, and come back when you're ready. Then you decide.
Step 4: Where should sessions be stored?
Store sessions in the database.
- How many people will be signed in at once.
- Whether you need to sign someone out from another device.
- What happens when the database is slow.
- How sign-out works in the app today.
- What your hosting allows.
- How the rest of the team handles this.
What breaks if I choose the cookie?
Step 4
A plan before any code, and a memory that lasts.
Underneath, Plinth plans first and codes second. The agent proposes a plan: the tasks, their order, and the files each task will touch. Nothing runs until you approve it, and each task is held to the files it was approved for. The work runs in a sealed space, and it has to prove itself against tests it isn't allowed to change, so it can't rewrite a test to make it pass.
The approved plan
Approved by you
Task 1
Task 2
Task 3
The sealed space
Your code
Plinth also keeps a memory of your codebase across sessions: what the code looks like now, what has been learned about it, and how it has changed over time. So it can tell you that the file you're about to touch broke something elsewhere the last time it changed.
Three kinds of memory
Now
Learned
Over time
Step 5
It started as a cathedral.
The first version of this idea was much bigger. I called it Cathedral: a whole system of roles and rules around every line an AI writes. The problem I saw was more than bad code. It was falling standards, lost focus, overloaded context, poor architecture, and confident mistakes. Making the AI smarter doesn't fix that. Governing how it works does.
That's still the deeper answer. But you don't start a cathedral with the spire. I decided to build the base first.
Plinth: first
Step 6
Two tools, pulling opposite ways, on purpose.
Keel is built for running many agents at once and seeing all of them. Plinth is built for sitting with one run and owning every choice in it. Most engineers need both, on different days. I'd rather build each one properly than blur them into one tool that does neither well.
Keel
Plinth
The AI can be fast. The decisions stay yours.
Share passwords and keys with your team without leaving them in chat forever.
Previous: TeamOS