Claude computer use vs Codex computer use vs ShadowCursor: a fair comparison
How Claude, Codex, ShadowCursor and the open-source Cua Driver and Peekaboo compare on background use on the Mac, Windows, several agents at once, clients, plans, clipboard and price.
Short answer: Claude and Codex both include computer use in their own apps, and on the Mac both can now work in the background. Claude's runs in the Claude Desktop app and the Claude Code CLI on Pro or Max plans; Codex's runs in the ChatGPT desktop app. If you work in one of those apps with one agent at a time, the built-in option is usually enough. ShadowCursor is for computer use in any MCP client, on plans the built-in features do not cover, and for several agents at once, each with its own cursor, keyboard and clipboard.
What is being compared
All of these tools let an AI model look at an app, click, type and read the result. They differ in where they run, which plan you need and what happens to your screen meanwhile. The comparison follows each vendor's documentation or README as of October 2026; computer use is still a research preview at Anthropic, so check the linked sources before relying on a single row.
New to agents working in windows behind yours? The guide to background computer use on the Mac explains the basics.
The comparison table
| Tool | Background on the Mac | Windows | Agents at once | Clients and plans | Clipboard | Price |
|---|---|---|---|---|---|---|
Claude Code CLI (computer-use server) | No: hides your other apps until the turn ends | No, macOS only | One session holds a lock | Claude Code CLI, Pro or Max; not Team, Enterprise, Bedrock, Vertex or Foundry; not in claude -p | Requested per session as an extra permission | Included in Pro or Max |
| Claude Desktop and Cowork | Yes, background mode; otherwise hides your windows | Yes | Not documented | Claude Desktop app, Pro or Max; not Team or Enterprise | Not documented | Included in Pro or Max |
| Codex in the ChatGPT desktop app | Yes, for a scoped task | Yes, but only in the foreground | Not documented | ChatGPT desktop app (Codex or ChatGPT Work), supported regions | Works with the target app's clipboard; separation not documented | Plans not listed on the docs page |
| ShadowCursor | Yes, with its own cursor, keyboard and windows | Beta | Several, each with its own cursor | Any MCP client: Claude Code, Codex CLI, Cursor, VS Code; also Claude Team, Enterprise and API keys | Its own clipboard per agent | 7-day free trial, then $12.99 per month or $69.99 per year |
| Cua Driver (open source) | When the app and platform support it | Yes, also Linux | README demo shows two sessions (on Linux) | CLI, MCP or SDKs (Python, TypeScript, Rust) | Not documented in the README | Free, MIT license |
| Peekaboo (open source) | For targeted input when it can resolve the process | No, macOS 15 or later | Not documented in the README | CLI, menu bar app, MCP clients | Not documented in the README | Free, MIT license |
Sources: Claude Code computer use, Claude Desktop, ChatGPT computer use, Cua README, Peekaboo README.
Claude computer use
Anthropic ships the same computer use engine in two places.
In the Claude Code CLI it is a built-in MCP server called computer-use that you enable with /mcp. The first time Claude needs an app in a session, you approve it. While Claude works, your other visible apps are hidden so it only interacts with the approved ones, and they come back when the turn ends. Only one session can use the computer at a time: the first session takes a lock and keeps it until it exits. It is built for testing your own apps from the terminal.
In Claude Desktop and Cowork computer use runs on macOS and Windows. On macOS it has a background mode (since September 2026) in which Claude works in the approved apps while you keep working. Outside that mode it hides your other windows.
Both share fixed control tiers by app category: browsers and trading platforms are view-only, terminals and IDEs are click-only, everything else gets full control. For browser work Anthropic points to Claude in Chrome instead. Both need a Pro or Max plan and a claude.ai login.
Codex computer use
OpenAI's computer use lives in the ChatGPT desktop app, where you select Codex or switch ChatGPT to Work, and install the Computer Use plugin. On macOS it can run a scoped task in the background while you keep working elsewhere. On Windows it runs on the active desktop: expect it to move the pointer, type and take over the foreground while the task runs. It asks before using each app, and it cannot automate terminal apps or ChatGPT itself. The documentation page does not say which ChatGPT plans include it.
ShadowCursor
ShadowCursor is a Mac app and MCP server. It does not replace your AI; it gives the agent you already use a cursor, keyboard, clipboard and windows of its own:
- Its own cursor, drawn in the agent's color with a name tag.
- Its own keyboard: keys arrive only in the agent's window, in one app.
- Its own clipboard: copy and paste never touch your clipboard or another agent's.
- Its own windows: it works in background windows, even when yours cover them, and screenshots capture only the window it works in.
Because it is an MCP server, it works in Claude Code, the Codex CLI, Cursor, VS Code and other MCP clients, including on Claude Team or Enterprise and with API keys. Several agents can work at once, each with its own cursor. The pointer moves in curves and typing has a human rhythm at about 200 words per minute, so hover menus (in Safari), sliders, date pickers and signature fields react as they do to a person.
Its limits, stated plainly: minimized and hidden windows cannot be operated; system drag and drop of files, tabs or text is cancelled rather than dropped; hover without a click does not reach Chrome or Electron apps; an agent opens its own browser window instead of using a background tab of yours; it needs macOS 15 or later, with Windows in beta. There is no free tier.
The open-source options
Cua Driver is part of the Cua repository and MIT licensed. It lets agents inspect and operate native apps and browsers on macOS, Windows and Linux through a CLI, MCP or typed SDKs. Its README says background delivery works without moving your pointer or taking focus "when the app and platform support it", and links a platform support page for the boundaries. The repository also offers Cua Spaces, full agent desktops in local VMs: a separate machine instead of windows on yours.
Peekaboo is an MIT-licensed macOS CLI and menu bar app for screen capture, accessibility inspection and UI automation, which it can expose to MCP clients. Targeted input uses background delivery when Peekaboo can resolve the process, so the app does not have to come to the front; some key chords require explicit foreground consent.
Both are free and scriptable. Neither README describes a per-agent cursor or a separate clipboard.
When to use which
- You use Claude Desktop on Pro or Max and run one task at a time. Use the built-in computer use with background mode. It is included in your plan.
- You test your own Mac app from the Claude Code CLI. The built-in
computer-useserver fits well, as long as hidden apps during a turn and one session at a time are fine for you. - You work in the ChatGPT desktop app on a Mac. Codex computer use in the background covers single tasks.
- You are on Windows. Claude Desktop runs there; Codex runs there in the foreground. ShadowCursor for Windows is still in beta.
- You want free, open-source tooling and are fine configuring it. Try Peekaboo on the Mac or Cua Driver across platforms, and test background behavior with the apps you actually use.
- You want computer use in the client you already work in, such as the Claude Code or Codex CLI, Cursor or VS Code, on any plan including Claude Team, Enterprise or an API key. That is what ShadowCursor is for.
- You want several agents at once, for example Claude on invoices and Codex testing a web shop, each with its own visible cursor and clipboard. ShadowCursor.
- You want an agent to operate websites, not just read them. Claude's built-in computer use keeps browsers view-only. ShadowCursor lets an agent work in a Safari window of its own with human-like pointer movement.
Whatever you choose, read whether it is safe to let an AI agent control your Mac before you approve apps.
FAQ
Is Claude computer use better than Codex computer use?
Neither is better across the board. Claude offers it in a CLI and a desktop app with per-app approval and fixed control tiers; Codex offers it in the ChatGPT desktop app and can operate browsers but not terminal apps. On the Mac both can work in the background. Pick the one that matches the app and plan you already use.
Can Claude and Codex use my Mac at the same time?
With the built-in tools, the Claude Code CLI allows only one session at a time, and neither vendor documents running several agents at once. With ShadowCursor, Claude, Codex and other MCP agents can work side by side, each with its own cursor, keyboard and clipboard.
Does Claude computer use work on Team or Enterprise plans?
No. According to Anthropic's documentation (October 2026), computer use requires Pro or Max and is not available on Team, Enterprise or third-party providers such as Bedrock. ShadowCursor works as an MCP server in any client, including Claude Team, Enterprise and API keys.
Is there a free alternative to ShadowCursor?
Yes: Cua Driver and Peekaboo are open source under the MIT license. The built-in options are included in their plans. ShadowCursor has a 7-day free trial without a card and no free tier.
Can Claude computer use control my browser?
Only view it. Claude's computer use treats browsers as view-only and points to Claude in Chrome for browser tasks. Codex computer use can work in browsers. With ShadowCursor an agent opens its own Safari window and clicks, types and fills forms there.