CLI Agent Orchestrator MCP Server

<!-- An absolute raw URL, not a repo-relative path: this file is also the PyPI long description (see `readme` in pyproject.toml) and PyPI cannot resolve relative paths. The banner is used rather than the bare mark because it supplies its own background, so the mark's navy never has to survive a backdrop it cannot see. The bare mark would need a <picture>/prefers-color-scheme swap, since that navy contrasts at only 1.77:1 on GitHub's dark theme, and PyPI allows <img> but not <picture>, so that markup would either be dropped or escaped into visible tag soup depending on how it sanitizes. The same constraint means the banner cannot be swapped per theme either. It is light, matching PyPI and GitHub's light theme, and it shows as a bright panel on GitHub's dark theme. That is a deliberate trade, not an oversight. Keep in sync with README.zh-CN.md. --> <p align="center"> <img src="https://raw.githubusercontent.com/awslabs/cli-agent-orchestrator/main/docusaurus/static/img/cao-social-card.png" alt="" width="640"> </p>

CLI Agent Orchestrator (CAO)

English | 简体中文

CLI Agent Orchestrator (CAO) coordinates multiple AI coding CLIs so a supervisor can delegate work to specialist agents in parallel or sequence.

📚 Documentation — guides, reference, and two interactive courses.

What CAO does

CAO runs a local cao-server, starts provider CLIs in isolated terminal sessions, and gives a supervisor tools for coordinating workers. The agents remain full CLI processes with their native authentication and capabilities. See CODEBASE.md for the runtime architecture and package layout.

Prerequisites

Install:

• Python 3.11 or later
• tmux 3.3 or later
• uv
• At least one supported provider CLI, authenticated before you launch CAO: Kiro CLI, Claude Code, Codex CLI, Antigravity CLI, Hermes, Kimi CLI, MiniMax Code, GitHub Copilot CLI, OpenCode CLI, Oh My Pi(OMP) CLI, Cursor CLI, or Grok Build CLI

The focused provider guides contain installation, authentication, and provider-specific behavior.

Install CAO

Install the current main branch as a uv tool:

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

For a tagged release, install cli-agent-orchestrator from PyPI. See DEVELOPMENT.md for a source checkout. For container-based installation, see the devcontainer feature.

To update an existing CAO installation:

cao update

See Updating CAO for source-aware behavior and edge cases.

First supervisor launch

The unqualified commands below use CAO's default Kiro CLI provider. If you installed a different provider, follow its focused guide above for the provider override while keeping the same sequence.

1. Install the built-in supervisor profile:

   cao install code_supervisor
   

2. In terminal A, start the local server and leave it running:

   cao-server
   

3. In terminal B, change to the project directory the agents should work in, then launch the supervisor:

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

4. Observe the supervisor in the attached launch terminal, open the Web UI at http://localhost:9889, or follow the tmux guide to attach to its session.

5. Stop the named session when finished:

   cao shutdown --session {session-name}
   

   To stop every CAO session instead, run cao shutdown --all.

Where to go next

Operate CAO

• Control-plane selection: choose the Web UI, shell CLI, operations MCP server, or plugins.
• Web UI and MCP Apps: browser and host-rendered fleet interfaces.
• Flows and workflows: scheduled runs and multi-step pipelines.
• Launch decisions: local routing settings and content-free records.
• Skills: install, scope, and author reusable agent guidance.
• Skills lifecycle example: install, discover, load, apply, update, and remove a reusable skill in an isolated home.
• Memory and self-learning: persistent cross-session memory, and the opt-in loop that turns workflow outcomes into lessons and promoted instructions.
• AI-DLC portfolio example: coordinate parallel AI-DLC intents across repositories and isolated worktrees.
• Cross-zone robot transport: coordinate simulated transport, measured custody handoffs, and independent checks.
• Hierarchical robot factory: a supervisor and three team-manager agents run nine Open Duck Mini robots on a simulated make, assemble, and ship line, and balance its bottleneck.
• Self-hosted decision model: run the open JevK5-4B decision model locally or on AWS (CDK), behind a loopback-only /v1/systemone API.
• Tool restrictions: roles, allowlists, and provider enforcement.
• Kubernetes deployment: run a supervisor and worker fleet on Amazon EKS, with shared workspace, per-pod state, and credential delivery.
• Updating CAO: update an installed uv tool.

Configure and integrate

• Ephemeral agents: opt-in, single-use helper agents that another agent creates and launches, and their limits.
• Ephemeral agents setup runbook: step-by-step setup of ephemeral agents and a decider, written for an agent to follow.
• Agent profiles: profile schema, discovery, provider selection, and overrides.
• HTTP API and PTY WebSocket: route-family overview and terminal streaming contract.
• Agent plugins: installing portable Agent Plugins 1.0.0 packages, and CAO's own operator and contributor packages. Every surface — the CLI, TUI, web panel, and the /plugins* HTTP routes — is built but gated off by default pending a naming decision. Set CAO_AGENT_PLUGINS_ENABLED to open it; until then the routes answer 404, exactly as an unregistered path would.
• Event plugins: outbound events, installation, and authoring.
• Provider behavior: Kiro CLI, Claude Code, Codex CLI, Antigravity CLI, Hermes, Kimi CLI, MiniMax Code, GitHub Copilot CLI, OpenCode CLI, Oh My Pi(OMP) CLI, Cursor CLI, and Grok Build CLI.
• Security policy: vulnerability reporting and deployment guidance.

Contribute

• Codebase guide: runtime surfaces, package ownership, and data flow.
• Development guide: local setup, testing, and verification.
• Release guide: maintainer release process.

Contributing

See CONTRIBUTING.md and DEVELOPMENT.md before submitting changes. Documentation changes must also follow the documentation maintenance rule.

License

This project is licensed under the Apache License 2.0. See LICENSE.