One window for every agent I have running, and one rail for everything waiting on me.
Ten agents in ten terminal panes means ten places to look before you can answer the only question that matters first thing in the morning: does anything need me?
Corral is my answer to that. It runs on the ranch server and puts every live agent conversation in one window — Claude Code beside Codex beside Grok beside a local model on the Mac mini — with every permission prompt they're blocked on collected in a single rail down the right-hand side.
It's a control plane, not a runtime. Corral never executes fleet work itself; it opens conversations, shows what they want to do, and carries my answer back. A selftest asserts the hub contains no subprocess, no os.system, and no dispatch route at any privilege level. The power stays with the agents, where their own gates can see it.
This is personal infrastructure. There's nothing to download and no repo to star; it knows my machines, my notes, and my habits, and it would be useless to you as-is.
The shape generalizes, though. If you run more than two agents at once you already have the problem Corral solves — work scattered across terminals, approvals discovered late, and nowhere that answers what's blocked right now. I'm publishing the screenshots because the pattern is worth arguing about, not because there's a signup at the end.
Corral speaks ACP, so which agent runs is a choice rather than a rewrite. Claude Code, ChatGPT's Codex, Grok, a local model on the Mac mini, a chartered cloud box, or a plain SSH shell — each gets a pane, and each pane wears the model, the effort, and the context it's actually running.
When an agent wants to edit a file or run a command, the prompt carries the diff, the exact tool arguments, and the file's hash before I decide. An unanswered prompt is a refusal, never a quiet yes — the agent stays as blocked as it was.
Cloud boxes, chartered workers and their expiry clocks, host heartbeats, what each box costs — read from what every one of those systems already publishes about itself. No live cloud API call happens on a page load, and a box with no recorded price says so instead of guessing.
Subscriptions, month-to-date API spend, and the projection against my monthly target. Corral computes none of it — every figure is another tool's own published number, printed verbatim. A missing cost renders as unreported, never as zero.
One search across the notes vault, the decision records, session briefs, research, and every project's own docs — about 1,200 pages, and each hit says which brain it came from. Read-only by construction: the hub has no write path into any of them.
The landing page, and the one I want to be boring: what fired overnight, what failed, what's still waiting. Looking like there's nothing to do is the win condition, so it has to be trustworthy when it says so.
Corral doesn't get powers the agents don't have.
Every pane is a real agent process on the host, with the same tool gate it would have in a terminal. Corral starts it, renders it, and gives that gate a button. That's the whole trick — and it's why the surface can be this open without being reckless.
One thing Corral does own is the permission posture. Panes deliberately don't inherit my ambient shell config — an early measurement caught that config silently suppressing every prompt, which would have produced a rail that was empty, faithful, and useless. So Corral writes its own config for each pane and shows the live posture as a pill on the pane's title bar.
One lane is missing on purpose. DeepSeek is third-party, and a pane isn't a classified API call — it's a shell in my working tree, next to my notes and my keys. There's no request to classify and nothing to refuse, so there's no lane. Gemini is listed but refused up front: the vendor retired the client, and a pane that opens and immediately dies is worse than an honest "no".
Live, on the ranch server. Real agents, really blocked — but on throwaway work, because my actual queue is nobody's business. Public addresses are masked and scratch paths shortened; nothing else is touched.
Corral is the newest piece of a personal AI setup I've been building and writing about for a while. AI-OS is the on-ramp, AI-OS Seed is the floor underneath it, and the notes are where I work out whether any of it survives contact with reality.
If you'd rather have the thinking than the screenshots, start with the notes — and tell me what you'd have built differently.
Comments ride GitHub Discussions — sign in with GitHub and say where I'm wrong.