From autocomplete to agents
My AI coding journey started like most developers: with GitHub Copilot autocomplete inside VS Code.
At first, it was just completion. A better autocomplete. Useful, but still part of the old way of building software.
Then I discovered Cursor.
Cursor made AI feel more integrated into the editor. It was no longer only suggesting the next line. It could understand more context, help with changes, and become part of the coding loop.
I used Cursor until Claude Code launched.
Claude Code was the first moment where agentic coding really clicked for me. I could give a task to an agent, let it explore the codebase, make changes, run commands, and come back with something I could review.
Making agentic coding programmable
At that point, I was using Anthropic models everywhere.
I used Claude Code heavily, and I started building my own tools around this workflow.
I built Spawner, a unified TypeScript interface to spawn and interact with AI coding CLIs like Claude Code, Codex, OpenCode, and Pi.
I also built Looper, a small engine to keep running an AI coding CLI against a prompt until it says it is done.
Basically, I was trying to make agentic coding more programmable.
I also started running long tasks on a VPS with Zellij.
The idea was simple: if an agent task needed to run for a long time, I didn't want it to die because my laptop went to sleep or because I closed the wrong terminal. So I used a remote server, persistent terminal sessions, and Looper to keep things running in the background.
It worked.
But it also made something obvious: I was slowly building a workspace out of terminals, scripts, sessions, and remote shells.
Why Pi became the center of my workflow
Then I discovered Pi.
Pi clicked with me because it felt different.
Not just another vendor CLI. More like a portable agent harness.
It is model-agnostic, extensible, and not locked to one provider's product, subscription, or master prompt. I can build on top of it without feeling trapped inside someone else's workflow.
That matters to me.
I want my agent workflow to move with me. If I use Anthropic, OpenAI, Kimi, or another model tomorrow, the workspace should not collapse. The harness should be a layer around the models and subscriptions I choose, not a cage.
I was already becoming a fan of Pi, but the real trigger came when Anthropic updated their terms around programmatic usage.
For a long time, this area felt vague. Then suddenly, the rules changed in a way that could break the kind of workflow I had built.
And that was the point for me.
They could break my workflow overnight. And they did.
So I cancelled my Claude subscription and decided to make Pi the center of my agent workflow instead. From there, I can use subscriptions or API access depending on each provider's terms, without tying my whole workflow to one vendor.
Around the same time, I was using Kimi 2.6 through OpenCode Go, and I started moving to a more portable setup.
Practically, I moved from paying around $100/month for Claude plus $10/month for OpenCode Go, to $20/month for OpenAI plus $10/month for OpenCode Go.
But the main point was not only saving money.
It was portability.
Cheaper was nice.
Sovereignty was the real reason.
The terminal wall
At that point, my daily workflow was mostly Pi inside Ghostty.
It worked well. I liked the simplicity of staying close to the terminal, and Pi gave me the portable agent harness I wanted.
Then I started seeing tools made to manage terminal sessions better, especially for AI workflows. That is how I found cmux.
So I tested it.
And honestly, cmux does a lot of things right.
It gives you session grouping.
It supports multiple tabs.
It works well with SSH.
It can keep terminal sessions organized.
It even gives you browser preview close to the agent workflow.
For a terminal-first workflow, it gets very close.
But I kept hitting the same pain point.
Sometimes I don't want another terminal pane.
Sometimes I want to edit code like I'm in an IDE.
Sometimes I want a native diff view.
Sometimes I want to inspect changes visually, review them properly, and make small manual edits without jumping back to VS Code or another editor.
That is where the terminal-first model breaks for me.
cmux helped me see the shape of the workspace I wanted, but it also made the missing piece obvious:
The agent workflow needs a real GUI layer.
Not only terminal sessions.
Not only panes.
Not only browser preview.
A real workspace with agents, code editing, diffs, browser preview, debugging, GitHub, and project context living together.
So I'm building Space Zero
Space Zero is a desktop workspace for solo builders using agentic workflows.
The vision is one place to build a product from idea to production-ready.
A place where you can capture an idea, turn it into a task, start an agent session, edit code, inspect your running app, debug issues, review diffs, generate assets, run tests, connect GitHub, open a PR, and keep project knowledge in the same workspace.
Not another chat app.
Not another terminal multiplexer.
Not just an AI IDE.
A real builder workspace.
Agents should not live in a disconnected terminal.
The browser should not live in another window.
Diffs should not require another tool.
Project knowledge should not be scattered across notes, docs, and old chats.
GitHub should not feel bolted on at the end.
Space Zero should bring all of this together:
- agent sessions
- code editing
- native diff review
- browser preview
- debugging context
- GitHub issues and PRs
- project knowledge
- testing
- asset generation
- voice/dictation for faster prompting
- eventually remote jobs for long-running work
The first loop I want to prove is simple:
Select something in your running app, tell an agent what to change, watch it work, review the diff, and preview the result — all inside Space Zero.
That loop is the seed.
But the larger goal is clear: a calm command center where solo builders can move from idea to shipped product without constantly losing context.
Space Zero is where you build products with agentic workflows, without friction.
I'm going to focus on this for the next six months.
If this sounds useful to you, join the waitlist:
Follow along on X: @spacezerodev