Vibe Coding in Xcode: Why I Built a Local AI Agent for Apple Developers 🧑💻
I found my productivity/ability to execute on my ideas jumped into warp with the announcement of Xcode 26.3 with Agentic Coding in February allowing some of the more-involved products I'd been wanting to tackle to come to fruition, whether it’s a simple sorting puzzle game for Apple Watch (which launched this week!) or getting the old Griffin Powermate Control Dial working on modern macOS versions (TestFlight available!) robust code assistance has been a game changer.
With this increase in productivity of app development, there’s also been some issues.
2026 Issues with Claude
Let me start by saying I’m bullish on Anthropic as the best AI Tool at the moment 💛 not only does it understand my prompts 9/10 times and deliver better results from those prompts, I also find its responses pleasant to read.
All that being said, 2026 has been a rough year from Claude. Everything from token usage exploding in March/April and there being some oddities with Opus 4.7 (at least, subjectively, for my use cases) just not being as effective as 4.6; or the whole nonsense with the smallest penis in the world; a lot of us have been having a love-hate relationship with Claude, especially since a lot of us have come to rely on it for our work.
The worst is still the 5-hour window. You’re getting into a flow and the Agent is popping all over the place and it just halts. My Apple Watch has practically become an “AI Timer” dedicated to counting down until my usage window resets.
Problems with Codex/ChatGPT
I’ll just make a small point here that the amount of free usage on Codex is insane, but it didn’t take too many times of it failing to create a markdown file before I switched back to Claude 🤦 Sorry, not sorry.
Apple Silicon Macs are Beasts 🐇
I’m currently running a 14-inch MacBook with an M2 Pro and I feel no current need to upgrade. The main reason I upgraded from my M1 Mac mini was portability alongside my Apple Vision Pro.
This got me thinking about how, with all this untapped power in our Macs, could a solution be made to at least supplement our frontier model usage if we hit a limit or it starts to burn tokens like they’re going out of style or, someday, replace it with an unlimited local source of coding assistance. No data centre required.
Local AI Tools are NOT user friendly 🐑
I’m not the smartest tool in the shed but I also know my way around computers, services, localports, etc. It’s obviously not impossible, but it can feel NON-obvious to get the right model for your specific hardware specs and getting it to consistently boot in a way that you can utilize.
AXIØM Vibe makes local AI coding assistance easy
Given that we’re using Apple Platforms and have certain expectations of how smooth and simple an experience should be, I thought I would make a contribution to help unlock the creativity of a new generation of vibecoders who are building for themselves. So I’ve built AXIØM Vibe, the easy local coding assistant that works right within Xcode.
An Actual Swift Coding Assistant
I have a very poor opinion on cross-platform frameworks like Flutter or React or Orgsm or whatever they’re called 🙄 Apple put it best in last year’s WWDC State of the Union:
40m48s–Discusses how native code is usually the better choice than cross-platform frameworks as, though seductive, they usually result in more complexity and code due to workarounds and recreating platform conventions, and are generally poorer at Accessibility, batter life, privacy, etc.
So my app spins up a local coding assistant that is:
a 7B coding model quantized to fit within your Mac’s memory, whether that’s
the previous base 8 gigabytes
the new standard of 16 gigabytes
or a loaded-up 32 gigabyte+ beast 🏎️
Optimized specifically on Swift code and Apple Frameworks
Can a 7B model compete with 500B+? The answer is “It Depends.”
Even though frontier models are incredibly impressive, and getting better every few months, these 500B+ behemoths have plenty of parameters and world knowledge that doesn’t apply to coding, let alone coding in Swift utilizing Apple’s frameworks and paradigms.
The Qwen2.5-Coder-7B model variant provides fantastic output while still being small enough to fit within average hardware and have usable inference timeframes. On its own, one still cannot be expected to live up to frontier models.
But then, my random boy Erik went and optimized it on >5k samples of clean Swift code, and that is where you start to get some very impressive results from a model that can fit in your MacBook Pro.
But what about the MacBook Air or the Mac mini?
The 16-bit version can barely load and provide stable inference on a 32 gigabyte Mac, but the majority of tinkerers are probably running base-level Macs which, up until recently, started with 8 gigabytes of memory.
Quantization does reduce model quality to some degree, but far less than many developers expect, and AXIØM Vibe has 5 different tiers designed to:
fit each of the different memory upgrade levels that have been offered from the M1-M5 releases of Apple Silicon Macs
provide a recommendation based on the user’s exact specs
load with a one-click download from Apple-hosted Background Assets that is lightning fast
Local Coding Assistance = Private Coding Assistance
Whether you’re working on sensitive projects or not, you can try flipping off Wi-Fi and then using AXIØM Vibe for coding and you’ll see it all occurs on-device. You do not have to trust me that your data isn’t used for training future models, your code never leaves your device with AXIØM Vibe.
The Future
Full agentic workflows are absolutely where I want to go next, but the current Mac App Store sandboxing requirements mean that I cannot access the Xcode MCP tools as they are currently structured (FYI I have submitted FB23099237 if any of y’all would also want to make that suggestion via Feedback Assistant).
Local AI Coding Assistant for Xcode launches right after WWDC 2026! 💻
You can Pre-Order the app now on the Mac App Store, and the full app will launch on June 16th, the Tuesday after WWDC!
The business model is “Pay-Once-for-LIFE!” Since you’re running the models locally on your Mac, and the model variants are made available through Apple-hosted Background Assets, I have very few ongoing costs other than working on what the next model will be and supporting macOS 27+ (and evaluating what new features could improve the experience and functionality).
My hope is that everyone leaves WWDC 2026 inspired to build something new, and that AXIØM Vibe can help turn those ideas into Swift code a little faster!
