Visualizing my swarm of coding agents on a live map

I’ve always felt like the standard single-chat interface for AI agents is kind of a trap. Don't get me wrong, it's fine for simple tasks, but once you start running multiple coding agents in parallel—like Claude Code and Codex—things get messy fast. I used to constantly lose track of which subagent was doing what, or worse, I’d realize thirty minutes too late that an agent had silently crashed or finished its task. It's a nightmare for anyone trying to manage complex workflows.

Since I already had all these JSONL transcripts being generated in my .claude or .codex directories, I figured, why not just build a way to actually see them? I put together a local web UI that runs entirely on localhost. It reads those existing logs and turns them into a live chat feed. The coolest part is the "parentage tree" I built; it scans the transcripts to reconstruct the entire lineage of the conversation, from the main root down to the tiniest background shell task. It basically gives you a visual map of your agent swarm.

The best part is that it doesn't mess with your existing setup. You can keep running everything headless in your terminal or tmux panes. The UI is just a viewer, but if you realize a subagent missed a tiny detail, you can actually send a message directly into its pane to fix it. I’ve also added a sidebar to track Claude Code and Codex usage limits, plus some voice transcription and workflow features.

I actually "vibecoded" the whole project using Claude Fable 5 and Codex. It was a massive test for the models—Codex struggled with the earlier iterations, but Fable absolutely nailed the logic.

I honestly think I built this because my ADHD makes keeping ten parallel agent threads in my head impossible, but I bet a lot of you feel the same way. If you're on Linux (or want to try the limited macOS support), give it a spin.

To try it out, just tell your agent: bunx agent-log-viewer and set it up.