Skip to main content

CLI Agent Orchestrator

CLI Agent Orchestrator (CAO) is a lightweight orchestration system for managing multiple AI agent sessions in tmux terminals. It enables hierarchical multi-agent collaboration through supervisor and worker agent patterns.

Why CAO?

  • Agent isolation — Each agent runs in its own tmux window with full context separation
  • Multiple orchestration patterns — Handoff (synchronous), Assign (asynchronous), Send Message (direct communication)
  • Multi-provider support — Works with 9 CLI agent providers
  • Scheduled flows — Cron-like task scheduling for automated workflows
  • MCP Server — Expose orchestration capabilities to any MCP-compatible client
  • Web UI — Monitor and manage agent sessions from a browser dashboard
  • Profile system — Discover and configure agent profiles dynamically

Supported Providers

ProviderIdentifier
Claude Codeclaude_code
Kiro CLIkiro_cli
Codexcodex
Kimi CLIkimi_cli
GitHub Copilot CLIcopilot_cli
OpenCode CLIopencode_cli
Hermeshermes
Cursor CLIcursor_cli
Antigravity CLIantigravity_cli

Note: Antigravity CLI is the Google-backed provider (formerly Gemini CLI).

Installation

uv tool install cli-agent-orchestrator

Quick Start

# Start the API server (required — run in a separate terminal)
cao-server

# Launch a session with a profile (uses your default provider)
cao launch --agents code_supervisor --session-name my-session

# Send a follow-up message to the session
cao session send cao-my-session "Start the code review workflow"

# Shut down when done
cao shutdown --all

A profile defines an agent's role, tools, and system prompt. Run cao profile list to see what's available. See the Quick Start guide for a full walkthrough.

Key Concepts

TermMeaning
SessionA tmux session containing one or more agent terminals. Created by cao launch.
TerminalA single agent running in a tmux window. Identified by an 8-character hex ID.
ConductorThe first terminal in a session — typically your supervisor agent.
ProfileA markdown file defining an agent's role, system prompt, provider, and tool access.
ProviderThe AI CLI backend (e.g., claude_code, kiro_cli) that powers an agent.
MCPModel Context Protocol — an open standard for connecting AI agents to tools. CAO exposes its orchestration as MCP tools so agents can spawn/message other agents.