March 25, 2026
Minimal code, maximal chaos
Building a coding agent in Swift from scratch
Fans cheer the DIY loop, lawyers side-eye the name, and Apple-local hype lights a fire
TLDR: A Swift project recreates a Claude-style coding assistant to prove simple tools and a tight loop work. Comments erupt over trademark worries, dreams of a fully local Apple-powered agent, and hard lessons about memory bloat—plus a library drop for shortcuts—showing the DIY agent race is heating up.
Someone just rebuilt a Claude Code-style coding helper in Swift, betting that fewer, better tools and a tight loop beat big, complicated frameworks. The 9-part learning series shows how a simple run-and-reply cycle plus file and search tools can carry most of the weight, with the AI doing the heavy lifting.
The crowd? Split and loud. Minimalists are applauding the “less scaffolding, more smarts” vibe, while the first splash of drama hit over the name: mark_l_watson warned, “don’t call it ‘claude’,” and everyone suddenly imagined lawyers stretching. Then came the platform war: one commenter asked if you could plug in Apple Intelligence for a “100% local” box, triggering debates about what’s possible on-device versus the cloud and whether mixing Apple and Google tech even makes sense.
Meanwhile, the realists showed up with receipts. maxbeech described long sessions drifting into “let me check that again” loops and mushy tool choices, pushing tips like summarizing finished steps to keep the bot sharp. A dev even dropped a shortcut: Operator, a Swift library for the agent loop. Meme lords chimed in with renames like “Claud-ish” and “Definitely Not Claude,” while Swift tinkerers formed a mini arms race to ship their own agents today.
Key Points
- •The project rebuilds a Claude Code-style coding agent in Swift to test minimal architecture principles.
- •A simple, invariant agent loop communicates with Anthropic’s Messages API and dispatches a small set of high-quality tools.
- •Hypotheses include thin orchestration, explicit task state, controlled context injection, and treating context compaction as a product feature.
- •A two-phase roadmap covers core loop mechanics (SPM setup, bash tool, file tool dispatch, todo reminders) and product features (subagents, skill loading, context compaction, task system, background tasks).
- •Architecture uses a two-target SPM setup (Core library and CLI), AsyncHTTPClient for raw HTTP, and runs on macOS and Linux; it is not intended as a full Claude Code clone or general-purpose framework.