Codex Computer Use Ventura
Restores OpenAI Codex Computer Use on macOS 13 by building a clean-room MCP backend that adapts to the bundled sky interface, enabling screen interaction and app control without modifying system libraries.
README
Codex Computer Use for macOS 13
An unofficial, source-built compatibility layer that restores the bundled Codex Computer Use skill on macOS 13 Ventura without modifying system libraries, lowering Mach-O deployment targets, or redistributing OpenAI code.
Tested on macOS 13.7.8 arm64, Xcode 15.2 / Swift 5.9.2, and bundled
Computer Use plugin versions 1.0.1000387 and 1.0.1000451.
[!IMPORTANT] This project is not affiliated with or endorsed by OpenAI. It does not contain OpenAI's proprietary Computer Use wrapper,
@oai/sky, skills, services, assets, or binaries. It interoperates with a legitimately installed local plugin and builds its native backend from MIT-licensed source.
Why this is needed
Recent bundled Computer Use services declare macOS 14.4 as their minimum OS
and link APIs unavailable on Ventura. Changing LSMinimumSystemVersion or the
Mach-O load command is not sufficient because the binaries genuinely import
newer Foundation and Swift runtime symbols.
This project takes a different route:
- Build an open Swift Accessibility/CoreGraphics backend targeting macOS 13.
- Adapt its stdio MCP tools to the ten-method
skyobject expected by the existing Codex skill. - Install a small clean-room loader in the local plugin cache. On macOS 14.4+ it delegates to the untouched saved OpenAI wrapper; on older systems it uses the open backend.
There is no patching of OpenAI native binaries and no TCC bypass.
flowchart LR
Skill["Bundled Computer Use skill"] --> Loader["Open clean-room loader"]
Loader -->|"macOS 14.4+"| Original["Saved local OpenAI wrapper"]
Loader -->|"macOS 13 / early 14"| Adapter["Sky-to-MCP adapter"]
Adapter --> Backend["MIT Swift macOS backend"]
Backend --> AX["Accessibility + CoreGraphics"]
What we added
- A Swift 5.9-safe capture result handoff for the open backend.
- The complete bundled Sky-facing tool surface, including
select_text. - Xdotool-style keypad names such as
KP_0. - Screenshot-pixel to global-coordinate conversion with bounds checks.
- App-path and bundle-ID resolution, app-scoped input, compact state diffs, permission prompting, bounded screenshot storage, and backend recovery.
- A hash-gated installer that refuses unfamiliar proprietary wrapper versions.
- An optional LaunchAgent that reapplies the loader when an update creates a fresh versioned plugin cache.
The backend source and history live in
Torrekie/open-codex-computer-use.
Requirements
- Apple Silicon Mac running macOS 13 or newer.
- Xcode 15.2 or another Swift 5.9 toolchain capable of targeting macOS 13.
- A current Codex/ChatGPT installation that already includes the bundled
Computer Use plugin and
node_replruntime. - Node.js 20+. The installer can use ChatGPT/Codex's bundled Node runtime.
- Accessibility and Screen Recording permission for the locally built helper.
Install
git clone https://github.com/Torrekie/codex-computer-use-ventura.git
cd codex-computer-use-ventura
./scripts/install.sh
The default build is ad-hoc signed. It works, but rebuilding can change its TCC identity and require granting permissions again. If you have a stable local code-signing identity, use it:
CODESIGN_IDENTITY='Apple Development: Your Name (TEAMID)' ./scripts/install.sh
The installer builds the pinned public backend commit, assembles a local app,
backs up recognized plugin wrappers byte-for-byte, installs the clean loader,
and enables the update watcher. Set INSTALL_WATCHER=0 to skip the watcher.
After installation, restart Codex or reset node_repl. The first app-state
request may open Privacy & Security; enable Accessibility and Screen
Recording for Codex Computer Use for macOS 13, then retry.
Verify
Use the normal bundled skill bootstrap. A minimal read-only probe is:
var apps = await sky.list_apps();
nodeRepl.write(JSON.stringify({ target: sky.target, count: apps.length }));
Then inspect Finder:
var state = await sky.get_app_state({ app: "Finder", disableDiff: true });
nodeRepl.write(JSON.stringify({
app: state.app,
textLength: state.text.length,
hasScreenshot: state.screenshot !== null,
}));
Updates
The watcher runs the hash-gated patcher whenever the bundled plugin cache changes. Known identical wrappers are updated automatically. Unknown future wrappers are left untouched and reported in:
~/.codex/computer-use-macos13/reapply.log
Do not add an unknown hash without reviewing the new wrapper and skill API.
Restore or uninstall
Restore the original local wrappers and disable the watcher while keeping the backend:
./scripts/restore.sh
Restore and remove the installed compatibility files:
./scripts/uninstall.sh
Neither command modifies macOS TCC databases.
Development
Run the Node adapter, loader, and installer tests:
npm test
Build and test the Swift backend fork:
BACKEND_SOURCE=/path/to/open-codex-computer-use ./scripts/build-backend.sh
See Architecture, Testing, and Troubleshooting for implementation details.
Security and licensing
Read SECURITY.md before enabling the backend on a sensitive machine. The bridge is MIT-licensed. The native backend retains its upstream MIT license and attribution; see Third-Party Notices.
OpenAI and Codex are trademarks of their respective owner.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.