Apple released Xcode 26.3 on February 26, bringing Anthropic's Claude Agent and OpenAI's Codex directly into its IDE as full coding agents. Greg Joswiak, Apple's SVP of Worldwide Marketing, confirmed the Mac App Store availability in a post on X Thursday.
This isn't autocomplete. The agents can explore your project's file structure, write code across multiple files, trigger builds, run tests, catch errors, and loop back to fix them. They can even capture Xcode Previews to visually verify that what they built actually looks right. Apple's newsroom post calls it "agentic coding," which is corporate-speak for: the AI does more than suggest, it acts.
Why it took so long
The release candidate first appeared on February 3. When iOS 26.3 and macOS 26.3 shipped on February 11, Xcode 26.3 was conspicuously absent. Apple pushed a second RC on February 20. Nobody at Apple explained what needed fixing between RC1 and RC2, which probably tells you something about how tricky it is to let autonomous AI agents loose inside an IDE that builds software for a billion devices.
"At Apple, our goal is to make tools that put industry-leading technologies directly in developers' hands so they can build the very best apps," said Susan Prescott, Apple's VP of Worldwide Developer Relations. Standard Apple press-release boilerplate, but Prescott's team did the actual work here: coordinating with both Anthropic and OpenAI to configure their respective agents for deep Xcode access.
The MCP angle matters more than the agents
Claude Agent and Codex get the headline, but the real story is buried a paragraph into every press release. Xcode 26.3 exposes its agent interface through the Model Context Protocol, an open standard that lets any compatible tool connect. Cursor, Claude Code CLI, custom internal agents: if it speaks MCP, it can access Xcode's 20 built-in tools.
Apple building on an open protocol instead of locking developers into two partners is a genuine surprise. The company has spent decades making Xcode a walled garden. MCP support means a team that prefers a different model, or has built proprietary tooling, now gets the same hooks that Claude and Codex use. One developer walkthrough confirmed all 20 native Xcode tools are accessible via MCP, including a semantic documentation search powered by Apple's internal MLX system (internally called "Squirrel MLX") that indexes Apple's entire docs corpus and WWDC transcripts.
There is a catch. Cursor currently chokes on some of Apple's MCP responses because Xcode sends non-compliant structured content. Claude Code and Codex handle it fine, which makes sense given they were co-designed with Apple. The workaround exists (a small Python wrapper), but the bug is on Apple's side.
So who actually uses this?
Early reception has been strong, at least among developers already comfortable with AI tooling. Steve Troughton-Smith posted threads on Mastodon about building a new app with minimal manual input and rewriting an entire Objective-C codebase to Swift using the agents. 9to5Mac noted that the response since the first RC has been "overwhelmingly positive," though it's worth remembering that RC testers are self-selecting enthusiasts, not a representative sample of the 30+ million registered Apple developers.
The more telling signal is what's happening in the tooling ecosystem around it. Projects like XcodeBuildMCP from Sentry already provide additional MCP tools for AI agents to control builds, debugging, and even physical device deployment. The infrastructure is growing faster than Apple's own release cycle.
Xcode 26.3 is available now on the Mac App Store and Apple's developer website. Xcode 26.4 is already in beta.
