APEX - AI Engineering OS

The first AI coding system that learns from your project, improves from your corrections, and gets measurably better every week you use it.

Quick Start · Why APEX · Commands · Intelligence · Stacks · Comparison

---

The problem every Claude Code user has

Every session starts from zero. You re-explain your stack. You repeat your conventions. You remind it that you don't use API routes. You paste your architecture again. The model forgets. You repeat.

By session 50 you've explained the same constraints 50 times. The model on session 50 is exactly as uninformed as it was on session 1. That's not a model problem. That's a systems problem.

APEX solves it.

---

⚡ Quick Start

Install:

cd your-project
unzip claude-orchestrator-apex-v4.1.zip
bash claude-orchestrator-apex-v4/install.sh

Then in Claude Code: one command, everything configured:

/setup

/setup detects your stack, auto-generates CLAUDE.md, seeds the project brain with your architectural constraints, and gets you to your first productive command in under 2 minutes: on any project, new or existing. No manual config required to start.

---

🧠 Why APEX

Every other Claude Code tool is stateless. Session 100 is identical to session 1. They add commands, they add agents, but they don't remember anything.

APEX compounds.

Session 1:  APEX knows your stack and hard rules
Session 5:  APEX has your first successful trajectory stored
Session 10: APEX knows your preferences for /design and /plan
Session 20: APEX injects months of real decisions before every command
Session 50: The gap between APEX and starting fresh is enormous

The underlying model doesn't change. Your accumulated project knowledge does.

---

🔧 The 18-Command System

Three groups. One mental model.

Meta / System

Command	What it does
/setup	Zero-friction onboarding. Auto-generates CLAUDE.md, seeds brain, warms cache. Works on any project.
/init	Start of every session. Validates context, syncs brain, confirms budget.
/status	Full system dashboard: brain health, cache stats, quality grades, DORA metrics.
/compact	Archive completed work, compress stale docs.
/benchmark	Statistical quality measurement for any command.

Dev Loop

Command	What it does
/brainstorm	Socratic requirements before any code. Generates Decision Record.
/ask	Read-only codebase query with brain context.
/plan	DAG-structured planning with trajectory injection.
/execute	Batched plan execution with lint+test between every step.
/design	Stack-adaptive UI with intentional aesthetic direction.
/spawn	Parallel agents in isolated git worktrees.
/test	Framework-specific test generation.
/debug	Root cause analysis.
/optimize	Performance profiling and targeted fixes.
/refactor	Safe refactoring with impact analysis.
/docs	Documentation generation.

Quality Gates

Command	What it does
/review	Multi-perspective deep review. Reads your AI_RULES.md, not just generic checks.
/ship	40-point pre-flight before any deploy. Runs your actual build and lint commands.
/rollback	Emergency rollback using worktree metadata.

---

🔬 The Intelligence Layer

Five Python modules that run locally, store persistently, and compound across sessions.

Project Brain

A temporal fact store with conflict detection. When you migrate from Prisma to Drizzle, the old fact gets invalidated and the new one takes its place automatically. No stale advice.

python3 .claude/intelligence/project_brain.py status
# Total facts: 19  |  Valid: 19  |  Constraints: 9  |  Patterns: 6

Trajectory Store

Based on NeurIPS 2025 research on self-generated in-context examples. Every successful session gets stored. When a similar task comes up, APEX injects your past wins before generating a single line of code.

python3 .claude/intelligence/trajectory_store.py query "build auth flow"
# → returns your previous auth implementation: what worked, what to avoid

Taste Memory

Explicit developer preference learning. After every /design and /plan, one question: was this on target? After 10 sessions, APEX knows you prefer functional components, dark backgrounds, queries in a central file. It stops suggesting the things you always change.

Evaluator

Deterministic self-scoring. Grades command quality from observable outcomes. Trends over time. Flags degradation before it becomes a habit. No additional API calls.

Semantic Plan Cache

Exact-match fast path plus synonym normalization. "Build auth flow" hits "create login system." 30–50% cost reduction at typical usage, growing over time.

---

🚀 Auto-Setup: No Config Required

The biggest friction in AI tooling is configuration. APEX removes it.

After install, run /setup in Claude Code:

✓ Language:   TypeScript
✓ Framework:  Next.js v15.2 (app-router)
✓ Database:   Supabase/PostgreSQL
✓ Commands:   npm run dev / build / lint

→ Generating CLAUDE.md...
→ Seeding brain with 7 architectural constraints...
→ Warming plan cache from TODO.md (12 templates)...
→ Ready. Run /brainstorm or /plan to start building.

Three things still need your input (takes 2 minutes): the project description, 2–3 project-specific conventions, and verifying your Hard Rules start with "Never" for brain sync. Everything else is automatic.

---

🏗️ Stack Support

Ecosystem	Frameworks
JavaScript / TypeScript	Next.js (App + Pages Router), React, Vue, Nuxt, Svelte, SvelteKit, Remix, Astro
Node.js Backend	Express, Fastify, Hono
Python	Django, FastAPI, Flask
Ruby	Rails, Sinatra
Go	Standard library + gin, echo, fiber, chi
PHP	Laravel, Symfony
Mobile	Flutter, React Native, SwiftUI

Version-pinned detection: APEX knows the difference between Next.js 13 and Next.js 15 because they are fundamentally different codebases.

---

📊 Honest Comparison

	Vanilla Claude Code	oh-my-claudecode	APEX
Persistent project memory	❌	❌	✅
Learns your preferences	❌	❌	✅
Trajectory replay	❌	❌	✅
Auto-generates CLAUDE.md	❌	❌	✅
Self-scoring quality system	❌	❌	✅
Version-pinned stack detection	❌	❌	✅
40-point pre-flight checklist	❌	❌	✅
DORA metrics tracking	❌	❌	✅
Multi-agent worktrees	❌	✅	✅
Gets better over time	❌	❌	✅
Windows Git Bash	❌	❌	✅
Python 3.8+ compatible	—	—	✅

The core difference: APEX compounds. Every other tool is stateless.

---

📦 What Gets Installed

.claude/
├── commands/          18 slash commands (including /setup)
├── intelligence/      11 Python modules: all local, no API calls
│   ├── generate_claude_md.py   Auto-generates CLAUDE.md from your project
│   ├── project_brain.py        Temporal fact store with conflict detection
│   ├── trajectory_store.py     Experience replay (NeurIPS 2025)
│   ├── taste_memory.py         Developer preference learning
│   ├── evaluator.py            Self-scoring quality engine
│   ├── benchmark.py            Statistical consistency measurement
│   ├── cache_manager.py        Semantic plan cache
│   ├── detect_stack.py         15+ frameworks, version-pinned
│   ├── token_tracker.py        Cost + time + DORA metrics
│   ├── design_system.py        Design token extraction
│   └── framework_lint.py       Framework-specific lint rules
├── references/        23 reference docs (stack-specific patterns)
├── scripts/           4 shell scripts (Windows-compatible)
├── config/            3 config files
├── brain/             facts.jsonl, grows with your project
└── memory/            trajectories/, taste signals, benchmarks

Nothing phones home. All intelligence runs locally via Python. No subscriptions. No accounts.

---

⚙️ Requirements

• Claude Code (claude.ai/code)
• Python 3.8+
• Git 2.20+
• Bash (macOS, Linux, Windows Git Bash, WSL2)

Optional: GitHub MCP (auto-PR in /ship), gitleaks (enhanced secret scanning)

---

🗺️ Roadmap

• v4.2 - Parallel /review subprocess harness (45s instead of 4min)
• v4.3 - Optional sentence-transformers for 85% semantic similarity
• v5.0 - Team intelligence: shared brain, averaged taste profiles
• v5.1 - CI/CD: APEX pre-flight as GitHub Actions gate on every PR

---

📄 License

MIT: use it, modify it, build on it.

---

Built by DevvNirvana

If APEX saves you time, a ⭐ helps other developers find it.