A month ago I wrote that I built a VS Code extension to work with several agents. The idea was to never leave the editor. A month later, Stack is no longer an extension.
What was happening
With several tasks open, VS Code, the extension and each agent's terminal added up to 1 GB or more. And my machine would grind to a halt.
The other thing is that inside VS Code I could not do it my way. An extension lives inside the editor: it has its sidebar, its panels and its commands, and it does not get out of there. I wanted a small window always on top, with my agents in view while I do something else. I wanted to jump from one agent to another with ⌃Tab, the way I switch between apps, and to have a shortcut take me to the one waiting for me from anywhere. You can't do that in an extension.
So I took it out. In a few hours, in one session, it was out of VS Code. The code viewer took a few more hours, in another. Two sessions and it was done.
Two pieces
stackd is what I already had: the extension's core, in TypeScript, now running as a background process. It keeps the tasks, receives the agents' hooks and serves the MCP tools. It talks to the app over a local socket.
The app is built with Tauri. When it opens it starts stackd, and if the one running is from another version, it replaces it.
The agents live in neither of the two. They are still in tmux, as in the extension. I can close the app or restart stackd and the agents don't notice.
What it uses
I measured it with the app open and my tasks running, using the same figure Activity Monitor shows:
| Measured | |
|---|---|
| Memory | Between 210 and 235 MB |
| CPU at rest | Around 1 % of one core, almost all of it stackd
|
Before, with VS Code, 1 GB or more.
But I wanted to read the code
That was the whole argument of the first post: agents write the code, but I have to read it.
So every task has a Code tab. It is read-only, built with Monaco, the editor inside VS Code: the files the agent changed, the tree, Quick Open, go to definition and find references, with PhpStorm's keys.
What mattered to me was go to definition in PHP. For that I use Phpactor, but indexing the whole repo in every worktree makes no sense when almost everything is the same in all of them. So there is one index per repository, of its base branch, built in the background. Each task copies that index and only re-indexes the PHP files it changed. For JS and TS there is a syntactic index plus git grep, and errors are flagged by oxlint and php -l.
It is not an IDE. It is just enough to read what an agent did without opening another app.
My way
This was what I could not do inside VS Code.
Now there are three modes: Full to see everything, Half to keep it in a column, and Float, that small window always on top. I move between agents with ⌃Tab, go to a task with ⌘1 to ⌘9, to the manager with ⌘0, and ⌘⇧J takes me to the next one that needs me. From any app, ⌃⌥⌘J brings me to the agent waiting for me.
A menu bar icon tells me how many agents are waiting, and from there I approve a permission without opening anything. That button also only answers if that same prompt is still on the agent's screen.
What I removed
I used the move to remove things I wasn't using or that didn't work the way I wanted.
Docker was the big one. Each task could bring up its own environment, but I can't have seven environments running at once. I'm working on another approach: a single Docker environment that switches to the task I'm looking at. It isn't there yet.
I also removed each task's debug port. In the first post I said I debugged with breakpoints. Right now I'm not debugging, so out it went.
And Codex, and the pull request status. Stack works with Claude Code and opencode.
What stays the same
Each task in its worktree, each agent in its tmux, a sidebar that tells me who needs me, the review with comments that go back to the agent, and the MCP manager that sets up tasks from a ticket.
All of this came from trial and error, and I don't know if it's the best way. But I no longer need a whole editor open to keep my agents in view.
The code is on GitHub, under the MIT license.
Top comments (0)