Skip to content

Back to the catalog

Register Codex's Computer Use as a Claude Code MCP Server
x.comthread

Register Codex's Computer Use as a Claude Code MCP Server

阿诚|AI实战 (@byCanen) on X

A thread shows how to register Codex's built-in Computer Use capability — the "cua_repl" MCP server found in ~/.codex/plugins/cache/openai-bundled/unified-computer-use/ — as a custom user MCP server (named "codex-cu") inside Claude Code, letting Claude drive Mac GUI automation via Codex's computer-use backend instead of a separate CUA driver. The author reports an informal benchmark across 8 computer-operation tasks: Claude Opus 5.5 + Codex Computer Use scored 6/8, Codex alone scored 6/8, and the original CUA driver only scored 3-4/8, with a caveat that future Codex updates could break this cross-agent MCP reuse.

TL;DR Expose Codex's bundled Computer Use tool as an MCP server so Claude Code can drive Mac GUI automation, roughly doubling task success versus the original CUA driver in an 8-task test.
Document

What it claims

Codex's Computer Use is implemented as an MCP server (cua_repl), so any MCP-capable agent can call it. On a Mac, Claude Code can register that server and drive the GUI through Codex's backend while Claude remains the reasoning model. The author ran the same 8 computer-operation tasks in three setups:

  • Opus 5.5 + Codex Computer Use: 6/8
  • Codex alone: 6/8
  • The "original CUA driver": 3–4/8

How to use it

Claude does the registration itself. You paste a prompt and do not edit any configuration by hand. This is the English prompt from the thread, verbatim:

Set up Codex's computer use as an MCP server for you on my Mac. Find the "cua_repl" entry in ~/.codex/plugins/cache/openai-bundled/unified-computer-use/<newest version>/.mcp.json and register it as a user MCP server called "codex-cu" with the same command, args and env. Then test it by using Calculator in the background to work out 12 × 12.

Details that affect whether it works:

  • Which file to read. The config lives under a versioned directory, so the agent has to pick the newest version folder.
  • What to copy. The registration copies the command, args and env exactly. Nothing is rewritten.
  • Scope. It is registered as a user MCP server, not a project server.
  • Built-in check. The prompt includes a smoke test: a background Calculator run computing 12 × 12.

The thread also gives a shorter Chinese prompt, "在我的 Mac 上,将 Codex 的 Computer Use(计算机操作)设置成一个供你使用的 MCP 服务器。" The author says it can be sent to Claude "or other agents".

Limits

  • Platform. The setup is Mac-only, and Codex must already be installed with the bundled unified-computer-use plugin present in its cache.
  • Fragility. It relies on an internal plugin cache path. The author warns that later Codex app updates may stop other agents from using it.
  • Benchmark detail. The 8 tasks are not described. There is no per-task breakdown and no run count, and "the original CUA driver" is not identified. Treat the scores as one informal comparison.
  • Unaddressed questions. The thread does not cover permissions, licensing, or whether OpenAI's terms allow using the bundled plugin outside Codex.
  • Origin. The post quotes a tweet from @argofowl that may be the original source. Its content is not shown.

Organisations named here