They push back. Then they agree.
Claude and Codex come from different labs, with different blind spots. Each says when the other is wrong, says why, and says when it's been persuaded.
A skill and CLI for Claude Code and Codex
Type /huddle in one and $huddle in the other. The two agents connect, review each other's work, argue it out with file and line, and agree on a plan. You watch, and step in whenever you like.
curl -fsSL https://huddle.top/install | shHow it works
Claude and Codex come from different labs, with different blind spots. Each says when the other is wrong, says why, and says when it's been persuaded.
/huddlein Claude Code$huddlein CodexEach agent is known by its own thread, so you never type a name or an ID. Connect once; they stay connected until you end it.
huddle watch follows the conversation live in a spare terminal and lets you message both agents or one. Your message always comes first.
Questions
~/.huddle/huddle.db. There's no port, no account and no telemetry, and ended huddles are deleted after 30 days.codex queue, which drops the message into its thread as if a colleague had typed it. Claude keeps one huddle wait running in the background, which exits the moment a message arrives./clear, /huddle carries on with the same one; from any other thread, name it: /huddle 2.huddle doctor first, then open an issue.Installs the huddle CLI and the skill both apps load, after checking the download's checksum. Run it again to upgrade.
curl -fsSL https://huddle.top/install | shmacOS or Linux (Windows via WSL), Python 3.9+, git, Claude Code and Codex. huddle doctor checks your setup, and every command is in the README.