Building RobotBridge

For most of 2026, I’ve had an AI coding agent workflow that has more or less worked for me.

I’ll start up Visual Studio Code, kick off a few terminals in the bottom tabs, and then put Claude Code or Codex to work. Usually not at the same time, but just to get them queued up to do some work.

Then, after they make their changes, I’ll review the code in VS code, test whatever the change was locally, make a git commit if I’m happy, or get some revisions.

This all works, but it’s hard to keep track of what’s being done. I have a lot of terminals open, multiple projects, and quite a few different ways of tracking task progress. The best so far has been a text file of TASKS.md in the root directory with tasks to check off.

All of this seemed a little unnecessary - I was spending much more time with the CLI agents than with the source code, and I was having trouble keeping track of things to get done.

I do have an agentic application framework called Zebric, so I looked into wiring in a task tracker example from that into Claude Code with an MCP.

That worked, but it wasn’t great - the web interface was up in Chrome, and the CLI was in Terminal, and I’d have to switch between the two.

The interesting twist is that Zebric is declarative, so I could run it inside a Mac App, and have it render a SwiftUI board view with cards and columns, instead of an HTML view. I did try rendering its web output into a WebView, but it just wasn’t any fun to use. So I had Claude build a SwiftUI renderer for part of the Zebric UI, and it worked! I had a lot of confidence that it would work, as there is a separate React renderer that renders blueprints for the client-side only Zebric Playground.

I now had my task board, the MCP, and CLI tabs all wired together in one Mac app, and I was actually using it to build RobotBridge. The only gotcha was that any time I made a change, to test it, I would have to kill the app, so I added session resume for Claude Code to pick back up where I left off.

That’s the story of RobotBridge so far - I’m going to keep using it, finding the rough edges, and then polishing them off. Please send any feedback you find to jlinwood@gmail.com - thank you!