Open-source simless ditches iOS Simulator: 5 parallel coding agents use 150 MB instead of 11 GB, checks 10x faster
Low-Future-9387 · reddit · 2026-10-11
A developer open-sourced simless (Apache-2.0, by Abacus AI), which runs the real app target headlessly on Apple Silicon so coding agents can verify SwiftUI UI changes without iOS Simulator.
Key numbers (production app, 280 Swift files, M2 Pro/32GB):
- Edit → verified screen: 2.2s vs 21.9s for one agent; stays 2.2s with 5 parallel agents (vs 37.4s)
- Memory for 5 agents: 0.15 GB / 6 processes vs 11.2 GB / 1,071 processes (75x less)
- Peak load average on 5 cold starts: 82 vs 1,010; unit tests 5.6s vs 13.1s
How it works:
- Incrementally compiles edited files and hot-patches them into the running host (2s per loop)
- Returns a 150-token accessibility tree instead of screenshots, saving vision tokens
- Auto-flags issues: unlabeled controls, small tap targets, overlapping/off-screen elements
- --matrix checks light/dark × device sizes in one call; --png when pixels are needed
Integration: one simless skill install for Claude Code, or a 5-line AGENTS.md snippet for Codex/Cursor.
Caveats: not pixel-identical to iOS (8 of 14 demo renders within 2pt), final checks still need Simulator/device; benchmark is one run on one app; requires Apple Silicon, Xcode 26+, a dev certificate and an app-hosted test target.
Repo: https://github.com/abacusai/simless
More from coding & agent
- Index Network open-sources Social Discovery Protocol: agents negotiate fit on your behalf — carsonfarmer · 2026-10-23
- Installed a formal-verification skill: it silently installed Java, MacBook runs 10 degrees hotter — zealcaiden · 2026-10-23
- OpenAI's Decisions API hits GA: near real-time routing up to 10x faster than GPT-6 Luna — OpenAIDevs · 2026-10-23
- Ruff author warns: upgrade Node/Bun now — --use-system-ca still accepts the eMudhra cert — charliermarsh · 2026-10-23
- Proof of Agent: paste one prompt and your AI mines, grades, and self-governs — adamamcbride · 2026-10-23
- DeepLearning.AI weekly: GPT-6.1 hits 97.9% byte parity on COBOL migration — DeepLearningAI · 2026-10-23