Detailed Analysis
A developer operating under the handle "greenpolo" has released an open-source Claude Code plugin called **multi-cli** that extends the architectural concept behind OpenAI's official Codex plugin for Claude Code — which allows users to delegate tasks to OpenAI's Codex CLI from within a Claude session — to encompass a broad range of competing CLI-based coding agents. The plugin, hosted at github.com/greenpolo/cc-multi-cli-plugin, ships with four pre-built subagent integrations: OpenAI Codex (`/codex:execute`), Google's Gemini CLI (`/gemini:research`), Cursor (`/cursor:debug`), and GitHub Copilot (`/copilot:plan` and `/copilot:research`). Beyond these defaults, the plugin includes a `/multi-cli-anything` skill that instructs Claude how to wire up any arbitrary CLI tool as a custom subagent, and a `/customize` skill that lets users reassign the roles of each integrated agent — for example, routing all writing tasks to Gemini or all debugging tasks to Cursor. The project is explicitly a community effort rather than an official Anthropic or OpenAI release.
The technical significance of the plugin lies in how it treats competing AI coding tools not as mutually exclusive environments but as composable, role-specific workers that Claude can orchestrate. OpenAI's official Codex plugin for Claude Code established the conceptual precedent: rather than forcing developers to context-switch between terminals, a session in Claude Code can spawn a Codex subprocess for targeted delegation. Greenpolo's multi-cli generalizes that pattern into a plugin framework with a consistent slash-command interface, effectively turning Claude into a meta-orchestrator. The `/multi-cli-anything` skill is particularly notable because it attempts to make the framework self-extending — Claude itself is instructed to locate an Agent Control Protocol (ACP) specification for a target CLI and configure the integration, rather than requiring the plugin author to manually hard-code each new tool.
The practical motivation the author cites — managing multiple free-tier AI subscriptions under tightening rate limits — reflects a real and growing pressure point in the developer tooling market. As commercial AI labs have moved to restrict third-party OAuth integrations and tighten API quotas, power users with access to several trial or free-tier accounts across different platforms find themselves unable to fully utilize those resources through conventional means. Multi-cli reframes this fragmentation as an asset: each subscription becomes a specialized worker with its own strengths. The author's framing resonates with a student or independent developer demographic that is simultaneously resource-constrained and technically sophisticated enough to route workloads across multiple agentic systems.
The plugin arrives at a moment when the broader industry is actively exploring multi-agent interoperability, but primarily through platform-layer solutions rather than plugin-layer ones. GitHub's Agent HQ, for instance, allows developers to choose between Claude, Codex, and Copilot within GitHub's own interface, but this approach is centrally managed by GitHub and constrained to its ecosystem. Multi-cli takes a decentralized, CLI-native approach — if a tool has a command-line interface and some form of machine-readable specification, it is theoretically integrable. This positions the project closer in spirit to Unix-style composability than to the curated agent marketplaces that larger platforms are building.
The project's limitations are worth noting: as a community plugin without official support, its reliability depends on the stability of each target CLI's interface and authentication mechanisms, both of which are subject to change at any vendor's discretion. The `/multi-cli-anything` skill's effectiveness also depends heavily on the quality of available ACP documentation for the target tool. Nevertheless, multi-cli represents an interesting grassroots response to the fragmentation of the AI coding agent landscape — an attempt to let Claude serve as a stable, user-controlled orchestration layer across a market that has not yet converged on a standard for cross-agent interoperability.
Read original article →