Noureddine RAMDI / MCO: orchestrating multiple AI coding agents through a neutral CLI layer

Created Mon, 04 May 2026 10:23:01 +0000 Modified Sat, 23 May 2026 20:41:27 +0000
998 Words
github-stars ai cli python multi-agent orchestration developer-tools

mco-org/mco

MCO (Multi-CLI Orchestrator) flips the usual AI coding assistant model on its head. Instead of relying on a single agent, it turns your AI IDE into a multi-agent review team by orchestrating several AI coding agent CLIs in parallel. This isn’t just running one AI tool after another; MCO dispatches prompts simultaneously, aggregates answers with cross-agent deduplication and consensus scoring, and can even synthesize results using an LLM. The neat trick is that the IDE’s own agent learns MCO’s CLI autonomously by reading its help output — creating a recursive orchestration where agents manage other agents through a neutral Python-based CLI layer.

what mco does and its architecture

At its core, MCO is a Python-based orchestration layer designed to run multiple AI coding agent CLIs like Claude Code, Codex CLI, Gemini CLI, OpenCode, Qwen Code, and OpenClaw in parallel. You provide a prompt, and MCO fans out this prompt to several agents simultaneously, collecting their outputs.

The architecture is command-line driven, with MCO acting as a neutral dispatcher that doesn’t depend on any single AI provider. It supports two main modes: review, which is optimized for structured code analysis and critique, and run, which handles more general AI coding tasks. The tool outputs results in various formats — JSON for programmatic consumption, SARIF for integration with code analysis tools, and Markdown for PR comments, making it versatile for different workflows.

Under the hood, MCO employs cross-agent deduplication to filter out redundant answers, and consensus scoring to highlight agreement across agents. It also supports optional LLM synthesis for summarizing or refining outputs.

MCO introduces several advanced modes to support complex workflows:

  • Debate mode: Agents critique each other’s outputs in rounds, enhancing the depth and quality of analysis.
  • Divide mode: Splits work among agents to cover more ground efficiently.
  • Chain mode: Enables sequential challenge workflows where one agent’s output becomes the next agent’s input.

Communication between agents can be structured through the ACP (Agent Communication Protocol) transport, and session memory persists across runs for context retention.

what sets mco apart technically

Several features distinguish MCO from typical AI coding assistants:

  • Neutral orchestration layer: Unlike tools tied to a single AI provider or platform, MCO sits above them all, enabling parallel dispatch to multiple independent AI coding agent CLIs.

  • Self-describing CLI: The CLI outputs grouped flags, defaults, and usage examples in terminal help (mco -h). This allows any AI agent capable of running shell commands to autonomously learn how to use MCO without additional documentation or training.

  • Cross-agent deduplication and consensus: MCO intelligently aggregates outputs by detecting duplicates and scoring consensus, which improves the reliability and relevance of combined results.

  • Multi-mode workflows: The debate, divide, and chain modes facilitate complex interactions and workflows between agents, supporting iterative critique and task splitting, useful for thorough code reviews or multi-step coding tasks.

  • Session memory: Persistent memory across sessions lets agents maintain context, enhancing long-running or multi-step workflows.

The tradeoff here is complexity: setting up and managing multiple AI coding CLIs requires some infrastructure and familiarity. Also, while MCO handles aggregation and orchestration, the quality and capabilities ultimately depend on the underlying AI agents used.

The codebase is clean and modular Python, with a focus on CLI UX and inter-agent communication protocols. The project favors extensibility and integration over user-facing GUI or IDE plugins.

quick start

Install via npm (requires Python 3 on your PATH):

npm i -g @tt-a1i/mco

Or install from source:

git clone https://github.com/mco-org/mco.git
cd mco
python3 -m pip install -e .

Run your first multi-agent review:

mco review \
  --repo . \
  --prompt "Review this repository for high-risk bugs and security issues." \
  --providers claude,codex,qwen

MCO’s CLI is designed to be agent-friendly and self-describing. Running mco -h or mco review -h shows grouped flags, defaults, and usage examples directly in the terminal, letting AI agents read and understand the interface autonomously.

This means you can tell your IDE’s built-in AI agent something like:

“Use mco to dispatch a security review to Claude and Codex, and a performance analysis to Gemini and Qwen — run them in parallel.”

The agent reads the help output, composes the commands, and orchestrates the workflow without additional prompts.

If you want to enable session memory, install the optional dependency:

pip install mco[memory]

verdict

MCO is a solid choice if you’re building multi-agent AI workflows, want to orchestrate several AI coding agents in parallel, or want to extend your AI IDE with multi-agent cooperation. The neutral CLI layer and self-describing interface are clever design choices that enable recursive agent orchestration and reduce integration friction.

That said, it requires managing the underlying AI CLIs, which can add operational overhead. It’s not a plug-and-play IDE plugin but a developer tool for orchestrating heterogeneous AI coding agents at scale.

For teams experimenting with multi-agent AI for code review, security analysis, or multi-step coding workflows, MCO provides a well-architected, flexible base to build on. It’s worth understanding even if you don’t adopt it directly, as the orchestration patterns and CLI design offer useful ideas for multi-agent AI integration.

→ GitHub Repo: mco-org/mco ⭐ 327 · Python