Most AI coding tools stop at writing code. They suggest a change, you paste it in, you build it yourself, and you find out whether it works. AgentSwift goes further. It is a native macOS app that runs an autonomous AI coding agent for Apple platform development: describe what you want to build, and it discovers your Xcode project, implements the change, compiles it, launches it in the simulator, and checks the result, without you touching Xcode. That closed loop, from request to verified build, is what makes it interesting.
Key Takeaways
AgentSwift is an open source, native macOS app that runs an autonomous Claude-powered coding agent for Xcode projects.
Its five phase workflow is Discover, Implement, Build, Launch and Validate, then Archive.
It builds with xcodebuildmcp, launches the app in a simulator or on macOS, and runs UI automation to verify behaviour.
You need macOS 26.1 or later, Xcode, Node.js, and an Anthropic API key. It supports Claude Opus 4.7 and Claude Sonnet 4.6.
It is an early project, with 126 stars and 40 commits at the time of writing, so expect rough edges alongside a genuinely useful idea.
Watch the 30-second repo signal
What AgentSwift Actually Is
AgentSwift, by hpennington, describes itself as a native macOS app that runs an autonomous AI coding agent for Apple platform development. You describe what you want to build, and AgentSwift uses Claude to discover your project, implement changes, build, run, and validate, without you touching Xcode. It is written in pure SwiftUI and Foundation, with no external Swift dependencies, according to its README.
That last detail matters more than it sounds. Many agent tools are web wrappers or Electron apps. AgentSwift is a Mac app built with Apple's own UI framework, which fits the workflow it is automating: it lives on the same machine as Xcode, the simulator, and your project, and it drives them directly.
The project is young. The repository was created on 2026-04-27 and, at the time of writing, has 126 stars, 40 commits, and 3 forks. Treat it as a promising early tool, not a finished product. Its value is in the workflow it demonstrates, and in how much of the tedious middle of development it takes off your hands.
How the Five Phase Workflow Works
AgentSwift drives a multi-phase agentic workflow, and each phase has a clear job. Understanding the phases helps you see where it saves time, and where you still need to stay in charge.
Phase | What happens |
|---|---|
Discover | Claude inspects your Xcode project structure and schemes |
Implement | The agent edits source files to match your request |
Build | It compiles the project using xcodebuildmcp |
Launch and Validate | It boots the app on a simulator or macOS and runs UI automation to verify behaviour |
Archive | The task is marked complete |
The first run discovers your project's scheme and simulator target. Later runs skip discovery and go straight to implementation, because the scheme, project path, and simulator ID are cached after the first build. That caching is a small detail with a big effect on everyday use: you only pay the setup cost once.
Two behaviours show the project was built by someone who has actually watched agents fail. First, message queuing: if you send a new message while the agent is running, the latest message supersedes earlier ones, so you can steer it mid-task instead of waiting for a bad run to finish. Second, error escalation: the agent attempts one fix on a failure, then surfaces the error to you rather than looping forever. An agent that knows when to stop is more useful than one that burns tokens retrying the same broken fix.
What You Need to Run It
AgentSwift is not a one click install, and the README is honest about that. The requirements are macOS 26.1 or later, Xcode, the Xcode command line tools, Node.js and npm, and an Anthropic API key. You also need two command line tools installed before the app will work.
Dependency | What it does | Install command |
|---|---|---|
xcodebuildmcp | Provides build, launch, and UI automation capabilities for Xcode projects | npm install -g xcodebuildmcp |
openspec | Tracks implementation specs across agent sessions | npm install -g @fission-ai/openspec |
Setup itself is simple once the dependencies are in place. Build and run the app in Xcode, open Settings and enter your Anthropic API key, select a project folder, optionally pick an iOS simulator from the dropdown, then type what you want to build and press Cmd and Return.
For models, the README lists two options: Claude Opus 4.7 for complex tasks and large codebases, and Claude Sonnet 4.6 for faster iteration and lighter tasks. That choice mirrors how developers already use Claude models elsewhere: reach for the stronger model when the change is large or risky, and the faster one when you are iterating.
Why the Timing Matters: Agentic Coding Comes to Xcode
AgentSwift did not arrive in a vacuum. Agentic coding inside Xcode has become a real, supported direction in 2026, which makes an independent open source take on the same idea worth watching.
Anthropic announced that developers can connect a Claude account to Xcode 26 for coding intelligence powered by Claude Sonnet 4. Then, in February 2026, Anthropic announced that Xcode 26.3 introduces a native integration with the Claude Agent SDK, the same underlying harness that powers Claude Code, including subagents, background tasks, and plugins, all without leaving the IDE. In that integration, Claude can be given a goal, break it down, decide which files to modify, iterate if something does not work, and capture Xcode Previews for visual verification.
Apple described Xcode 26.3 the same way: agents can search documentation, explore file structures, update project settings, and verify work visually by capturing Xcode Previews and iterating through builds and fixes. Xcode 26.3 also exposes its capabilities through MCP, an open standard.
AgentSwift is a different answer to the same problem. Instead of working inside Xcode, it works alongside it, as its own app, with its own five phase loop and its own choice of Claude models. That independence is both its appeal and its risk: it can move faster than Apple's release cycle, but it also has to keep its own build and automation tooling working as Xcode changes.
The Tools Behind the Agent
Two dependencies do most of the heavy lifting, and both are worth knowing on their own.
The first is xcodebuildmcp, the tool AgentSwift uses to compile, launch, and automate. Its repository, now renamed MobileBuildMCP, is maintained by Sentry and has 6,526 stars. It provides tools for agent use when working on iOS and macOS projects, and it ships as a single package with a CLI mode and an MCP server mode. Its README lists macOS 14.5 or later, Xcode 16.x or later, and Node.js 18.x or later as requirements. Note that opening the old XcodeBuildMCP URL now resolves to the renamed MobileBuildMCP repository, while AgentSwift's README still uses the old name and install command, so check the current instructions before installing.
The second is OpenSpec, which tracks implementation specs across agent sessions. The Fission-AI project has 73,484 stars and describes spec-driven development for AI coding assistants: each change gets a folder with a proposal, specs, design, and tasks, and it works with more than 30 AI assistants via slash commands. In AgentSwift, that means the agent's plan is not just a chat transcript that disappears. It is a spec that persists across sessions, which should make follow-up changes more consistent.
Together, these two tools explain the design. One gives the agent hands on Xcode, the other gives it a memory of what it planned to build.
Who Should Try It, and Who Should Wait
AgentSwift makes the most sense for Apple developers who already use Claude and want to automate the loop of writing, building, and checking a change. If you spend your day switching between an AI chat window, Xcode, and the simulator, a tool that runs that whole loop for you is easy to justify trying.
It is also a good fit for small, well scoped tasks: add a screen, fix a layout bug, wire up a new view. Those are the tasks where discovery is cheap, the build is fast, and simulator validation actually proves something.
You should probably wait, or at least keep expectations modest, if you need a stable daily driver, if your project has a complicated build setup, or if you are uncomfortable handing an agent an API key and build access to your project. With 126 stars and 40 commits, AgentSwift is early. Early open source tools change fast, break occasionally, and sometimes change direction. Try it on a side project first, and keep your main project on a workflow you already trust.
If you work in a different corner of the AI coding world, the same site has a related spotlight on tracking exactly which lines your AI agent wrote: Git AI: Track Exactly Which Lines Your AI Coding Agent Wrote. Pairing a tool that writes code with a tool that audits it is a sensible combination.
Frequently Asked Questions
Is AgentSwift free?
Yes, it is open source, but running it is not free in practice. You need your own Anthropic API key, and Claude API usage is billed by Anthropic. The app itself costs nothing; the model calls do not.
Does AgentSwift replace Xcode?
No. It works alongside Xcode. You still need Xcode installed, and AgentSwift uses it, through xcodebuildmcp, to build and run your project. It replaces the manual loop of editing, building, and testing, not the IDE itself.
Which Claude models does it support?
The README lists Claude Opus 4.7 for complex tasks and large codebases, and Claude Sonnet 4.6 for faster iteration and lighter tasks. You pick the model in the app after entering your API key.
What happens when a build fails?
The agent attempts one fix, then surfaces the error to you rather than looping. That error escalation is deliberate: it stops the agent from burning time and tokens on a fix it cannot make.
Do I need to know Swift to use it?
You need an Xcode project to work on, so some Swift or Apple platform knowledge helps, especially when reviewing what the agent changed. But the point of the tool is that you describe what you want in plain language, and the agent handles the editing, building, and validation steps.