If you’ve been doing a lot of agentic development, here’s a situation that probably sounds familiar. You’ve got 5 tabs open, working on 5 different projects, running Claude Code, Codex, maybe OpenCode.
As the coding agents improve, they work for longer, so naturally you multitask, getting another project moving forward while the agents crank away on the others. The problem is that it’s easy to lose track of what’s going on in the other tabs, and you end up clicking through them one-by-one to see where they are and whether they’re waiting on input. Then you come across one that finished whatever you asked four minutes ago and has been sitting there waiting for an answer to its next question. You don’t notice, because every tab looks exactly the same. To make it worse, it’s hard to keep straight which project is running in which tab.
I know it sounds like a focus problem, but when I talk to others doing agentic development, the stories are similar.
There are various ways to solve this problem. Good tools exist, but I do a lot of work directly in Linux, and I couldn’t find anything there that fit my needs, so I decided to build a terminal that works the way I do and can help keep track of what’s going on during working sessions.
Terminals are blind on purpose
Modern terminal emulators are impressive: GPU-accelerated, beautiful font rendering, fast. But they render glyphs. They don’t know anything about what’s happening inside them. To the terminal, an agent thinking hard for two minutes and an agent that stopped and is waiting for your input look identical: a process attached to a pty that happens not to be printing anything right now.
That was completely fine for fifty years, because you were the one typing every command. You always knew the state of your own session, because you were driving it. The moment you start delegating work to agents that run for minutes, even an hour or more, and then quietly wait, the terminal’s blindness becomes the bottleneck.
What vibeflow does
vibeflow is a Linux terminal emulator with one core idea: every tab carries a small indicator that tells you what the program inside it is doing. It shows the state two ways at once: a thin colored stripe down the tab’s left edge, and the state spelled out in the label.
Codex working (blue) between Claude and OpenCode waiting (amber). The amber tabs are the ones that need you.
The stripe’s color is the at-a-glance signal:
- Blue: the tool inside is working (generating, running commands).
- Amber, gently pulsing: it’s waiting on you. This is the one that matters.
- Gray: idle at a prompt; a normal running command shows nothing special.
That’s it. You glance at the tab bar and you know, at a distance, which conversation needs you back, without alt-tabbing through all of them to check. The amber tab is the one to click.
How it knows (the short version)
There are two ways vibeflow figures out a tab’s state, and they stack.
The clean way is an open escape sequence, which I call OSC 1338, that a tool can emit to announce its own state: “I’m working,” “I’m waiting.” Claude Code, for example, can be wired up with a few hooks so it emits these as it goes. Any tool can adopt it; it’s a published, documented protocol, not a private feature.
The fallback, for tools that don’t emit anything, is heuristic: vibeflow looks at which process is in the foreground of each tab and watches its output. When a known AI tool goes quiet for a few seconds after working, that’s almost always “waiting on you,” and the tab goes amber.
The first way is precise; the second works on day one with zero cooperation from the tool. If you want the actual wire format, the detection timing, and the parts that were genuinely hard to get right, plus the bugs daily use and a pre-launch audit surfaced, I wrote a deep-dive on how it works.
The Build
vibeflow is free and open source (MIT/Apache-2.0), GPU-rendered in Rust, and runs on Linux under both X11 and Wayland. It’s v0.1.7: early, useful daily for me, and rough in the places early software is rough.
I should also be upfront: I built it as one person, I’ve done a lot of work in Python, but I’m new to Rust, and I leaned heavily on AI agents (Claude Code) to write it. That felt fitting for a tool whose entire reason to exist is to help streamline agentic development, but it also means I’d welcome more eyes on it.
Try it
cargo install vibeflow
…or grab the single-file AppImage from the releases page. Source and docs are on GitHub.
If you build terminal tools or AI coding tools, I’d especially love for you to look at the OSC 1338 protocol. The whole point of making it an open escape sequence instead of a private feature is that it gets more useful the more tools and more terminals speak it.