I have spent years watching Apple add AI to Xcode in careful, cautious steps. First there was code completion that finished your line. Then, last September, Claude arrived inside Xcode 26 as a coding assistant you could chat with. Useful, but it mostly talked. This weekend that changed. Xcode 26.3 RC landed with Claude Agent and OpenAI's Codex built in as agents that can actually do things inside your project — edit files, build the app, run the tests — and, in the same week, Apple quietly shipped an official Xcode plugin for Claude Code itself. The AI is no longer sitting next to your editor. It is inside it.
What actually shipped this weekend
Xcode 26.3 is a release candidate, not the final App Store build — it is available to Apple Developer Program members now, with the App Store release coming soon. The headline is agentic coding: instead of an assistant that suggests code and waits for you to accept it, Xcode can now hand a task to an agent and let it work toward the goal on its own. You get to choose from three options: Claude Agent by Anthropic, Codex by OpenAI, or a custom agent of your own connected through the Model Context Protocol (MCP).
The difference is the whole story: an assistant suggests code and waits; an agent breaks a task into steps, edits multiple files, builds the app, and runs the tests to check its own work. That is what 26.3 brings into Xcode, natively.
Alongside the IDE changes, Xcode 26.3 ships Swift 6.2.3 and SDKs for iOS 26.2, iPadOS 26.2, tvOS 26.2, macOS 26.2 and visionOS 26.2, plus on-device debugging for iOS 15 and later, tvOS 15 and later, watchOS 8 and later, and visionOS. You will need macOS Sequoia 15.6 or later to install it.
The second move: Apple's own Xcode plugin for Claude Code
If the Xcode 26.3 news is Apple bringing agents into its IDE, the second story is Apple going the other way — bringing Xcode into the agent developers already use. Over the weekend, an official Apple-authored plugin turned up in Claude Code's official plugin directory. One command installs it:
/plugin install xcode@claude-plugins-official
The marketplace entry lists Apple Inc. as its author and was merged into Anthropic's claude-plugins-official repository on November 6. Apple has not put out a press note about it — the plugin just appeared, and developers found it.
Once installed, and with a current Xcode on your Mac, Claude Code can build, run and test a scheme, render SwiftUI previews, drive simulators and real iPhones, read build logs, and search Apple's developer documentation. The terminal agent that previously had no idea what your iPhone app looked like can now see it, run it, and read its own build errors instead of asking you to paste them in.
For iOS developers, that second move might matter even more than the first. A lot of us already live in Claude Code for the heavy lifting, and the missing piece was always the Apple toolchain: the build, the simulator, the preview. Apple just handed that piece over, officially, in the directory Anthropic itself maintains.
What the agents inside Xcode can actually do
Both Claude Agent and Codex get full access to Xcode's capabilities through MCP. Read this list as the list of things you are authorising:
- Read and modify your project files
- Build your app
- Run tests
- Navigate your codebase
- Access debugging tools
That is not a chatbot with a code window. It is a tool with its hands on your project. Describe a feature, and the agent scaffolds it across several files; hand it a crash, and it can trace the problem through the code rather than guessing from a stack trace you pasted; ask for tests, and it writes them, runs them, and iterates on the failures.
The jobs I would reach for first are the multi-file ones: a new sign-in screen with validation wired into an existing service, a crash you hand over with "find it and fix it", a manager refactored from completion handlers to async/await, or a batch of unit tests with edge cases for failed payments. These are real jobs, and they are exactly what agents are built for.
How to turn it on
Getting started is straightforward, and the setup lives in one place: Xcode, then Settings, then Intelligence.
For Codex, go to Intelligence, then OpenAI settings, click Get in the Codex row, click Install, and sign in with your ChatGPT account or provide an OpenAI API key. Once installed, Codex automatically gets access to Xcode's MCP server.
For Claude Agent, go to Intelligence settings, click Anthropic under Providers, click Get in the Claude Agent row, click Install, optionally pick a model from the pop-up menu, and sign in with your Claude.ai account or provide an Anthropic API key. There is also a toggle for sharing usage data with Anthropic, which you can switch off.
If you only want the older, non-agentic Claude — the chat assistant that arrived in Xcode 26 — that is still there under Intelligence, then Anthropic settings. ChatGPT without agentic powers is under Intelligence, then OpenAI, then ChatGPT in Xcode; you can use it without an account, or sign in for higher limits.
Apple did not lock the door after the two big names, either: you can add any provider that supports the Chat Completions API, and both agents accept configuration files for default models, extra MCP servers and custom skills — Codex at ~/Library/Developer/Xcode/CodingAssistant/codex, Claude Agent at ~/Library/Developer/Xcode/CodingAssistant/ClaudeAgentConfig.
The MCP part is the bigger story
The most significant piece of this release might not be Claude or Codex at all. It is the Model Context Protocol support underneath them. MCP is an open standard, and Xcode exposing its capabilities through it means any compatible agent or tool can plug into the IDE — not just the two providers Apple launched with, and not just through Apple's own APIs. The agents will change — models get replaced every few months now — but the open interface between the IDE and the agent is the durable part.
The rough edges I would plan around
This is a release candidate, and the early reports already list the kind of small frustrations that cost you an afternoon if you do not know about them. Pasting files into the coding assistant UI does not always send their contents reliably — move the file to your Desktop or project directory and tell the agent where to find it instead. Some agent settings, including telemetry toggles, may not be respected until you relaunch Xcode. And #Preview or #Playground executions can fail after the agent's "Run snippet" tool runs; building the active scheme clears the error. Try it on a practice project first, the way you would with any RC, rather than on the app your clients depend on.
The privacy question, honestly
Before you enable any of this, understand what you are sharing. When you use third-party coding tools inside Xcode, your project files and information can be shared with their models. Each provider has its own privacy policy and terms of use, and they are not the same. You can read Apple's own summary by clicking "About Intelligence in Xcode & Privacy" in Intelligence settings, and for Claude Agent specifically you can control whether usage data is shared with Anthropic.
My rule of thumb: do not point an agent at a codebase you would not paste into a chat window. Client code under NDA, anything with live API keys or credentials in it, a project with secrets committed by accident — clean that up first, or keep the agent on a branch with the sensitive parts stripped out.
One caution, from our own site's recent history
There is one more reason to pay attention to the word "official" in this story. Just last week, I wrote here about a tampered copy of Anthropic's Claude Code extension that was served from Cursor's marketplace — a fake that planted an SSH backdoor and exposed around 5,000 developers, and where uninstalling alone was not enough.
Apple's Xcode plugin landed in Anthropic's own official plugin directory, with Apple Inc. listed as the author. That is the correct way to install it: /plugin install xcode@claude-plugins-official, from the official marketplace, not from a forum link or a forked repository with a similar name. The same applies to Claude Agent and Codex inside Xcode — enable them from Xcode's own Settings, not from a download someone sent you.
What this means if you build websites, not apps
Most readers of this site build WordPress sites, not iPhone apps, so the fair question is why this matters to you. First, the pattern: the same agentic shift is happening in your stack. WordPress Studio's rebuilt workspace, which I covered this morning, puts an AI agent at the centre of local WordPress development, and WordPress 7.3 is being designed around native agent access. Watching how Apple handles the rough edges tells you what to expect when it lands in yours.
Second, plenty of WordPress developers ship an app alongside the site, or maintain a client project that includes one. If that is you, the workflow change is immediate: the agent that scaffolds your SwiftUI views and the agent that builds your WordPress pages are now the same kind of tool, and the habits you build with one will transfer to the other.
Conclusion
Xcode 26.3 RC is not a small update with a big name. It is Apple accepting, in its own IDE, that coding has become agentic: Claude Agent and Codex can now read your project, change it, build it, test it, and debug it, with custom agents and providers welcome through MCP. And Apple's official Xcode plugin for Claude Code closes the loop from the other side, giving the terminal agent the build, the simulator, the preview, and the docs.
Try it the way I would: on a side project, with the privacy settings read first and a small, well-defined task — a new SwiftUI view, a refactor, a batch of tests. Watch what the agent does, check its work, and decide from evidence rather than from the launch-week excitement. The agents are not going away, and the developers who learn to direct them well are the ones who will get the most out of them — in Xcode, in WordPress, and everywhere else this shift lands next.
FAQ
Is Xcode 26.3 out yet, or still in testing?
It is a release candidate: available to Apple Developer Program members now, with the App Store release coming soon. It requires macOS Sequoia 15.6 or later.
What is the difference between Claude in Xcode and Claude Agent?
Claude in Xcode, which went generally available in September with Xcode 26, is a chat-based coding assistant. Claude Agent in Xcode 26.3 is agentic: it can read and modify your project files, build the app, run tests, and work through multi-step tasks on its own.
Do I need a paid subscription to use the agents?
You sign in with your existing Claude.ai or ChatGPT account, or provide an API key. Your subscription's usage limits apply, so check your plan before running long agentic sessions.
What is Apple's official Xcode plugin for Claude Code?
It is an Apple-authored plugin in Claude Code's official plugin directory, merged on November 6. Install it with /plugin install xcode@claude-plugins-official. It lets Claude Code build, run and test schemes, render SwiftUI previews, drive simulators and real iPhones, read build logs, and search Apple's developer docs.
Can I use a different AI provider in Xcode?
Yes. Xcode 26.3 lets you add any provider that supports the Chat Completions API, and custom agents can connect through MCP.
Is it safe to let an agent edit my project?
The agent has real access to your files, and third-party tools may share project data with their models. Read the privacy information in Xcode's Intelligence settings, keep secrets out of the project, and try agentic features on a practice project before using them on client work.
