My AI assistants can ask each other for a second opinion. I can see the whole exchange, and answer their permission requests, in one window.
Updated · What it does now · Get Corral Light
August 2026 screenshot. The capabilities below reflect the October update.
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 started by putting those conversations and their permission requests in one window. Now I can ask several models the same question, have them review each other's answers, and come back to unfinished work without rebuilding the conversation.
Named roles save me from writing the same review instructions each time. Scheduled conversations, reminders, and phone notifications help me keep track of what needs attention; the fleet, spending, and notes are still in the same place.
The full Corral knows my machines, my notes, and my habits. It remains personal infrastructure. Corral Light is the public, MIT-licensed version: the conversations, shared permission rail, file search, exchanges between assistants, and a git branch of its own for each agent, without my ranch-specific rooms.
Light runs locally with assistants you have installed and signed in. Its setup guide starts with AI-OS Seed, the workspace underneath it. Both versions now share the same conversation and permission core, so a fix to that core reaches both.
Corral Light in two and a half minutes: the shared window, nine ways the assistants work together, and what sits underneath. No sound; the captions carry it. Every conversation in it is made up.
Full Corral on the ranch, plus two Corral Light features marked In Light. The Light guide describes what ships in the public version.
Claude Code, Codex, Grok, Gemini through Antigravity, and local Ollama models run side by side. Fireworks and direct DeepSeek add paid model options for work within their data limits. Cloud workers and SSH shells have their own panes; the picker shows which lanes are available.
Permission requests arrive in one rail with the command or tool arguments and, where available, the file diff. The decision is bound to the displayed request. Each assistant keeps its own permission rules; a waiting request never becomes an approval because I left it alone.
Ask several assistants at once, quote one answer into another conversation, or give each the others' answers to review. Agents can also request a consultation through the same window. The prompts and replies stay visible, and subscription-backed lanes use the logins I already have.
I can name a pane @author and another @reviewer. The author sends its work across, can wait for the answer, and can broadcast to every named pane. A message arrives marked as another agent's, never as mine. After four hops without me, the hub pauses the loop and asks.
An agent that needs my decision asks for it with a tool instead of ending a reply with a question. The pane reads needs you until I answer, and the question survives a restart.
Save the named panes as a rig and bring all of them back with one command. Each seat reports what happened to it: resumed, rebuilt, or started fresh.
In Corral Light an agent can work on its own git branch, in its own worktree, so my checkout is never touched. I review the diff, then commit, push, open a pull request, or discard with a recovery copy. With Review at the end, on Claude for now, edits inside the branch skip the card and the review is the approval.
Challenge this change hands the frozen diff and my acceptance criteria to a model from another vendor, never the author's reasoning. It reads a read-only copy in a sandbox, and its findings land on the lines. It approves nothing; the commit is still mine.
Pick a reviewer, an auditor, or an adversarial evaluator instead of rewriting their instructions. A role supplies the brief and defaults for a conversation, consultation, or scheduled job. It grants no extra permissions, and tells me when a lane cannot apply a requested setting.
Schedule a new conversation, a reminder, or a return to an existing pane. Resume a disconnected or dead pane when its adapter supports it. Local terminal sessions can be watched for liveness too, with expiring watches; watching one does not give Corral control of it.
For work governed by a runbook, a turn-end check looks for the required run and recorded progress. If the check refuses completion, the pane is parked with the reason in the rail. I can see what is missing before treating the work as finished.
Cloud boxes, host heartbeats, and chartered workers sit beside their costs and expiry clocks. Fleet views now include AWS, Google Cloud, and Hetzner. A signed charter waiting for a box is shown separately from a running worker; available shells and remote desktops open from the same window.
Subscriptions, month-to-date API spend, and the projection against my monthly target, alongside the spending tools' own signals. A missing cost renders as unreported, never as zero. Subscription-backed consultations and metered API lanes remain different costs.
Search the notes vault, decision records, session briefs, research, and project docs together. Each hit names its source. Select a passage and ask an assistant about it without losing the page you were reading.
Today brings forward blocked work and recent runs. Selected items can notify my phone through Home Assistant when no browser is watching, with quiet hours and a morning digest. The permission rail also reaches my terminal workspace, where I can read and answer a request.
Corral doesn't get powers the agents don't have.
Corral opens real assistant processes and carries their permission requests back to me. Passing a question to another model, selecting a role, or scheduling a conversation does not grant it more authority.
Permission controls differ by assistant. Corral can impose its own posture on Claude; other lanes manage permissions through their own tools. A role cannot make that distinction disappear. The interface reports a requested setting that cannot be applied instead of showing it as protection I have.
The model choices have changed since the first screenshots. Gemini now connects through Antigravity's native ACP server. Fireworks and direct DeepSeek are available for public or internal work; sensitive material remains outside those lanes. The subscription lanes and paid API lanes are labeled separately.
These screenshots were published August 27, 2026. They show the earlier layout, before the newer roles and consultation controls. Real agents on throwaway work; public addresses are masked and scratch paths shortened.
Five lanes, one window tells the story of the second opinions. The lanes learned to talk is the shorter account. Nobody Was Lying follows the work that led to the runbook gate. Give every agent its own branch covers own branches and agents handing work to each other.
To try the shared window on your own machine, start with Corral Light's setup guide. Let me know what it gets right — and where it still gets in your way.
Comments ride GitHub Discussions — sign in with GitHub and say where I'm wrong.