Xcode 26.3 adds AI agents

Xcode 26.3 ships with Swift 6.2.3 and SDKs for iOS 26.2, iPadOS 26.2, tvOS 26.2, macOS 26.2 and visionOS 26.2, and the release notes state a macOS Sequoia 15.6 requirement for running Xcode 26.3. ( developer.apple.com ) Apple documents that agentic coding in Xcode 26.3 is delivered via first‑party integrations with Anthropic’s Claude Agent and OpenAI’s Codex and is exposed through the Model Context Protocol (MCP). ( developer.apple.com )( developer.apple.com/videos/play/tech-talks/111428/ ) Xcode includes a built‑in MCP server bridged by xcrun mcpbridge that exposes roughly 20 native Xcode tools — examples surfaced by early analysts include BuildProject, RenderPreview (SwiftUI preview capture), DocumentationSearch, test execution and file/project operations. ( rudrank.com )( github.com/farkasseb/xcrun-mcpbridge-tools-list ) Enabling external agents is controlled in Xcode’s Intelligence settings, and Apple documents one‑line setup commands for external MCP clients such as 'claude mcp add --transport stdio xcode -- xcrun mcpbridge' and 'codex mcp add xcode -- xcrun mcpbridge' for immediate MCP connectivity. ( developer.apple.com ) The Xcode 26.3 release adds a comprehensive permission model and UI that lets users approve or restrict what agentic tools can do, including toggles for integrated internet access and granular tool approvals called out in the release notes. ( developer.apple.com ) Apple’s release artifacts for 26.3 also record updated agent binaries and compatibility notes — the update adds support for Claude 4.6 and upgrades the bundled Codex client to version 0.98.0. ( developer.apple.com ) Early community audits and guides have cataloged edge cases: researchers found mcpbridge responses that use an unstructured 'content' field rather than spec‑required 'structuredContent', and published gists captured agent launch invocations that include a '--dangerously-skip-permissions' flag used in experimental setups. ( dev.to )( )