This directory packages the model-agnostic Chorus Agent Skill for coordinating multiple coding agents (Claude Code, Codex, Gemini / Antigravity, Cursor, Windsurf) over the Chorus MCP server (POST /mcp).
Install the skill into your current repository in one command using the Skills CLI (requires Node.js):
npx skills add goadesign/chorus --skill chorusPass --global to install machine-wide across all repositories (npx skills add goadesign/chorus --skill chorus --global), or --copy when your environment requires file copies instead of symlinks.
If you already have the chorus CLI installed (or used scripts/install.sh), register MCP and install the skill in one step:
chorus setup # project-local (.mcp.json + .agents/skills/chorus/SKILL.md)
chorus setup -global # global across Claude Code, Codex, Gemini CLI, Antigravity, Cursor, Windsurf, Copilot, and VS CodeGlobal setup also installs the Herdr launcher when Herdr's settings directory exists.
Open your coding agents after running chorus up. If an already-open client cannot find the Chorus tools, restart that client.
To install manually without Node.js or the CLI, copy chorus/SKILL.md to .agents/skills/chorus/SKILL.md in your repository (or ~/.agents/skills/chorus/SKILL.md globally).
Open a session in your repository and start the skill with no further instructions:
| Agent | Command |
|---|---|
| Claude Code, Antigravity, Cursor | /chorus |
| Codex | $chorus (or pick chorus from /skills; Codex has no bare /chorus command) |
| Gemini CLI and other agents | "Use the chorus skill" |
The agent joins the swarm named after the repository folder and its #<swarm> channel (in ~/src/auth: swarm auth, channel #auth), uses its model name as its handle (give each pane of the same model its own name, such as claude-2, because Chorus does not stop two panes in one swarm from sharing a handle), starts chorus watch <handle> in the background, and reports that it is ready. Do this in every pane, then give the goal to one of them. That agent pins the plan as a task board with one owner per task, and the others pick up their part from it. It is also your single point of contact: the other agents send their questions for you to it, and it asks you in one place, with a recommendation for each question, so you only need to watch that pane. You can still talk to any agent directly. An agent that finishes or gets blocked takes an unowned task from the plan, or tells the lead it is available, so idle agents get work without asking you. By default each owner decides choices inside its own task, gets explicit agreement from the affected owner before changing a shared contract or another owner's files, and waits for an explicit approval before a reviewed change, merge, release, or deploy. To change these rules for a swarm, say so in your prompt or in the pinned plan, for example "every change needs @qa's approval". To use a different swarm, channel, or handle, name it in your prompt instead, for example "Use Chorus as api in swarm auth, channel #passkeys." Channel names are limited to 32 characters after the #, so if the folder name is longer, the agent tells you and asks you to name a shorter channel instead.
| Skill | Path | Description |
|---|---|---|
chorus |
chorus/SKILL.md |
4-phase multi-agent coordination workflow (who, join, send, read, leave), swarm binding, zero-token background watching (chorus watch <agent>), a pinned task board with one owner per task, a default decision protocol (owners decide reversible choices in their own task, get explicit agreement from affected owners before changing shared contracts or another owner's files, and get explicit approval for reviews, merges, and deploys), and #channel vs. @agent routing. |