codex

cc-multi-cli-plugin

If you have access to multiple AI coding CLIs (Codex, Cursor, Antigravity, and OpenCode), this plugin lets Claude Code delegate to whichever one is best for the task — without you having to switch tools or run them yourself.

Each CLI is wired up through its native transport (Codex via ASP, Cursor via headless agent -p, Antigravity via its headless agy CLI, OpenCode via headless opencode run --format json). This lets you pick and choose the best features from each — like /cursor:delegate for fast implementation, /codex:review for code review, /antigravity:research for deep research, or /opencode:delegate to offload implementation to OpenCode's Zen models. Sessions, streaming, tool calls, and background jobs all work normally.

Install

Paste into Claude Code:

/plugin marketplace add https://github.com/greenpolo/cc-multi-cli-plugin
/plugin install multi@cc-multi-cli-plugin
/multi:setup

/multi:setup detects which CLIs you have, installs the matching sub-plugins, and wires Exa + Context7 MCPs into each.

Skills Included:

Two skills ship with the plugin:

• multi-cli-anything — adds ANY CLI beyond the four built-in providers (Aider, Qwen, or anything with a headless/print, app-server, or structured transport) as a subagent that Claude can invoke at will. Claude scaffolds the new plugin in the marketplace.

• customize — change which CLI handles what. "Make Codex the executor instead of Cursor." Claude does the file edits, reinstalls, and tells you what restarts are needed.

Just ask Claude in plain English. The skills activate automatically.

Commands

Provider commands live under each CLI's namespace; the cross-cutting /multi:* commands operate the shared runtime.

Command	What it does
/codex:execute	Delegate a specific plan or plan step to Codex
/codex:rescue	Hand a stuck or open-ended problem to Codex for an independent investigation
/codex:review	Codex code review of your working tree or a branch (read-only)
/codex:adversarial-review	Adversarial design/code review — challenges the approach, not just the diff (read-only)
/cursor:delegate	Delegate an implementation task or plan step to Cursor (agentic; writes code; supports --until-done)
/cursor:research	Read-only external web/documentation research via Cursor
/cursor:explore	Read-only codebase exploration via Cursor
/antigravity:research	Deep external research with Antigravity (Gemini 3.5 Flash, read-only; experimental)
/antigravity:explore	Fast codebase exploration with Antigravity (Gemini 3.5 Flash, read-only; experimental)
/opencode:delegate	Delegate an implementation task to OpenCode (agentic; writes code; supports --until-done; default model: opencode/claude-opus-4-8 via Zen)
/opencode:research	Read-only external web/documentation research via OpenCode
/opencode:explore	Read-only codebase exploration via OpenCode
/multi:setup	One-shot wizard — detects CLIs, configures Exa + Context7 MCPs
/multi:status	Show active and recent background jobs for this repo
/multi:result	Show the stored final output for a finished job
/multi:cancel	Cancel an active background job

Claude can also auto-dispatch the provider commands without you typing them.

All of them are interchangeable, and can be altered to whatever you want using the customize skill.

Known issues

These are upstream CLI quirks and current limitations. If you hit something not listed, check the companion's stderr (the forwarders append 2>&1) — a bad model id, an auth failure, or a sandbox block surfaces there. (ACP_TRACE=1 still exists for ACP-based CLIs, but none of the shipped providers use ACP: Cursor runs headless agent -p and Antigravity headless agy.)

• Cursor runs in headless agent -p mode (not ACP). The adapter delivers the prompt on stdin, selects the model with --model (default auto), and parses json/stream-json output. This sidesteps the older ACP-mode bugs (silent MCP drop, model-hint quirks). MCP servers come from Cursor's own ~/.cursor/mcp.json, which /multi:setup maintains.