Exody: a macOS coding workbench with a 5-stage agent loop and per-step multi-model router
Smokiezzz · reddit · 2026-11-09
A developer unveiled Exody, a native macOS workbench for autonomous software engineering, positioning it against Cursor, Claude Code, and OpenAI Codex: no single-vendor lock-in, first-class support for local Ollama models and BYOK API keys with zero markup.
Key architecture:
- 5-stage autonomous loop: Investigate → Plan → Build → Verify → Fresh Diff Review. Tasks run in isolated Git worktrees with parallel sub-agents executing test suites, tsc, and headless browser/iOS Simulator checks, followed by an independent cross-model diff review before merging. Author claims 10/12 blind benchmark wins over raw frontier calls.
- Per-step Smart Model Router: picks between Claude, GPT, Gemini, Grok, DeepSeek, Kimi, or local Ollama per step based on task difficulty, tool-use accuracy, and token cost.
- Design Studio: measures typography and spacing, renders and critiques UI before shipping (6/6 blind visual wins), covering Web, WebGL 3D, and motion film.
- EXO Desktop Assistant + iPhone companion: local capture of Zoom/Meet/Teams audio, native Mac app automation, and remote run monitoring/diff approval over LAN or P2P WebRTC.
The author invites questions on routing heuristics, worktree orchestration, and cross-model review.
More from coding & agent
- Dev slams Codex app: cloud-to-local handoff is 'completely broken' — adhamelarabawy · 2026-11-09
- Scott Hanselman open-sources sandbox-mcp: disposable Windows VMs for AI agents — davemccollough · 2026-11-09
- PM runs 5-6 Claude Code agents in parallel for 24+ hours with 5 no-code principles — PawelHuryn · 2026-11-09
- Dev lets three Claude agents live in a virtual town; adding personalities makes them far more interesting — johnnypen2445 · 2026-11-09
- A single AI prompt turns agents into a product design review team that screenshots and fixes UI issues — vista8 · 2026-11-09
- Stanford and DeepMind evolve stale benchmarks into hard tests with TURTLE, targeting 50% solver success — rohanpaul_ai · 2026-11-09