LAC — Local Agentic Coding (Apple Silicon Macs)

Target: Apple Silicon Macs (8GB+ supported; best on high-RAM Studio + TB5 external).
Primary Model: Qwen 3.8 27B dense hybrid VLM with native MTP speculative decoding (~45 tok/s).
Harness: OpenCode V2 (@coder, @reviewer, @orchestrator).
Orchestrator: Hermes + Kanban Loops (~/todo/lac-loops/).
Gateway: lac-router on :8000 (auto-routes MLX :8080, llama-server :8081, Ollama :11434).

---

10x Pro Quickstart

# 1. System diagnostics & status
./rust-src/target/release/lac doctor
./rust-src/target/release/lac status

# 2. Start the unified gateway on :8000
./rust-src/target/release/lac route --daemon

# 3. Start high-speed inference (MLX with native MTP)
./rust-src/target/release/lac serve mlx

# 4. Or launch the interactive Terminal UI (TUI)
./rust-src/target/release/lac tui

# 5. Run OpenCode against the unified local gateway
opencode2 run "Verify stack: list files and check git status"

---

Architecture & Layout

• SwiftUI/ — LAC Studio native macOS desktop app (Apple Silicon native):
  • Apple Liquid Glass: Optically accurate specular refraction borders, ambient light sheen, dual-layer visionOS depth shadows, 120Hz ProMotion interpolating springs, and trackpad tactile haptics.
  • Claude & Codex Ergonomics: 780pt centered golden-ratio reading column, sparkle monogram avatar, syntax-highlighted code containers with animated copy buttons, collapsible sidebar (⌘B), + New Chat (⌘N), and floating elevated bottom composer.
  • Code Assistant (⌘2): Split coding canvas with multi-language editor on the left and agentic assistant on the right. Quick actions for Explain, Refactor, Generate Tests, Audit Bugs, and Optimize, with 1-click apply and undo.
  • LM Studio Hugging Face Hub: Live debounced search across 100,000+ HF models, Apple Silicon / MLX / GGUF filter capsules, unified memory RAM fit estimation, and 1-click model switching.
  • Live Ops Dashboard: Real-time daemon monitoring, one-click backend switches, port health telemetry, and self-healing error recovery.
• rust-src/ — 100% Pure Native Rust Backend (std library only, zero external crates, zero Python/shell scripts in the backend path):
  • lac — Swiss-army-knife CLI (status, serve, route, code, doctor, loop, bench, kv)
  • lac-router — Unbreakable intelligent reverse proxy on :8000 (SIGPIPE-immune, thread panic-isolated, auto-routes MLX :8080, llama :8081, Ollama :11434)
  • lac-tui — Full-featured interactive ANSI terminal control center
  • bootstrap — Idempotent machine setup and volume provisioning
  • serve-mlx — MLX-LM server launcher with port drift detection and MTP on :8080
  • serve-llama — llama-server with Flash Attention & Q8_0 on :8081
  • pull-models — Automated Hugging Face model puller with RAM fit gating
  • kv-manage — Context cache risk analyzer & truncation engine
• opencode.jsonc — OpenCode V2 project config (gateway :8000, 64K context, 14 skills, qwen3.8-27b primary)
• templates/ — Production Kanban queue (lac-tasks.yaml) & loops (daily-coding, bug-fix, feature-branch, security-audit)
• .opencode/skills/ — 14 skills including hermes-orchestrator, loop-orchestrator, kv-cache-manage, thermal-monitor
• launchd/ — KeepAlive daemons for lac-router (org.lac.router.plist) and serve-mlx (org.lac.serve-mlx.plist)
• docs/ — Technical blueprints (ARCHITECTURE.md, MTP_TUNING.md, LOOPS_AND_KANBAN.md)

---

Documentation

• Operating Rules (AGENTS.md)
• Quickstart Guide (QUICKSTART.md)
• Architecture Blueprint (docs/ARCHITECTURE.md)
• Speculative Decoding & MTP (docs/MTP_TUNING.md)
• Kanban & Autonomous Loops (docs/LOOPS_AND_KANBAN.md)
• Terminal Selection Guide (TERMINALS.md)