Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

awslabs/cli-agent-orchestrator

Developer Tools

Orchestrate multiple AI coding CLIs from one supervisor, with each worker running in its own tmux session.

1,592
336
Python
Apache-2.0
1 year old

Overview

CLI Agent Orchestrator (CAO) coordinates AI coding CLIs such as Claude Code, Kiro, Codex, Copilot CLI, and Cursor CLI so that a supervisor agent can assign work to specialist worker agents in parallel or in sequence. Each worker remains a real CLI process with its own native authentication and capabilities; CAO starts a local cao-server, launches provider CLIs in isolated terminal sessions (tmux), and gives the supervisor tools to manage them. It targets developers who already use one or more coding CLIs and want to split tasks across several agents without manually juggling terminals.

The documentation site covers guides, reference material, and two interactive courses; the codebase guide describes the runtime architecture and package layout.

Install

Prerequisites: Python 3.11+, tmux 3.3+, uv, and at least one supported provider CLI that is already authenticated. The README lists Kiro CLI, Claude Code, Codex CLI, Antigravity CLI, Hermes, Kimi CLI, Qwen Code, MiniMax Code, GitHub Copilot CLI, OpenCode CLI, Oh My Pi, Cursor CLI, and Grok Build CLI, each with its own provider guide; see for example the Claude Code guide and Codex CLI guide.

Install the current main branch as a uv tool, then check the CLI:

uv tool install git+https://github.com/awslabs/cli-agent-orchestrator.git@main --upgrade
cao --help

Tagged releases are available on PyPI; for a source checkout see the development guide, and for containers see the devcontainer feature. Update an existing install with cao update.

First supervisor run

The default provider is Kiro CLI. Install the built-in supervisor profile, start the server, then launch the supervisor from the project directory agents should work in:

cao install code_supervisor

In terminal A, keep the server running:

cao-server

In terminal B:

cd /path/to/your/project
cao launch --agents code_supervisor

Observe the supervisor in the launch terminal, open the Web UI at http://localhost:9889 (see the Web UI guide), or follow the tmux guide to attach to its session. Stop a session with cao shutdown --session {session-name}, or stop everything with cao shutdown --all. If you installed a different provider, use its guide's provider override while keeping this sequence.

Operating and extending

Beyond the basic supervisor, CAO documents flows and workflows for scheduled and multi-step pipelines, agent profiles for provider selection and overrides, ephemeral agents for single-use helpers, skills and memory for reusable guidance and cross-session state, and tool restrictions for role-based allowlists. A Web UI and API and an EKS deployment example cover browser control and cluster runs. Agent plugins are built but gated off by default behind CAO_AGENT_PLUGINS_ENABLED; until enabled, those routes return 404.

Constraints

CAO is a local orchestration layer, not a model provider: you need an authenticated provider CLI, and behavior differs by provider. It relies on tmux sessions, so a Unix-like environment with tmux 3.3+ is effectively required. The default provider in unqualified commands is Kiro CLI. Agent plugins are disabled by default. The project is Apache-2.0 licensed and its security policy documents vulnerability reporting and deployment guidance.

HotRepo's take

On 2026-11-19 CAO recorded 9 new stars for the day and a best position of 13 on the Python daily list; no all-language daily rank has been recorded, so its standing is only visible within the Python list. The project is squarely in the multi-agent coding workspace space, but its distinguishing detail is that workers stay full CLI processes under tmux rather than being reimplemented inside a framework, which keeps each provider's own auth and capabilities intact. The breadth of provider integrations and the provider-specific behavior differences are the main thing to evaluate before adopting it.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. getpaseo/paseoOrchestrate multiple coding agent CLIs across desktop, mobile, web, and terminal from one self-hosted daemon.TypeScript23,361
  2. smtg-ai/claude-squadA terminal TUI that runs multiple AI coding agents in isolated tmux and git worktree workspaces.Go8,789
  3. eneskirca/nodetermOrganizes terminals and parallel AI coding agent sessions as persistent, draggable nodes on an infinite canvas, so you can see everything at once.TypeScript2,287
  4. mvschwarz/openrigCoordinates Claude Code, Codex and other coding agents as a persistent team with roles, shared state and owned work.TypeScript8,246
  5. Yeachan-Heo/oh-my-claudecodeRun a coordinated team of Claude Code agents on one coding task instead of prompting a single agent repeatedly.TypeScript40,339