The ECC Config System: All 8 Modules Explained
Everything Claude Code's 8 core modules broken down — CLAUDE.md, Rules, Agents, Skills, Hooks, MCP, Commands, and Memory.
Everything Claude Code (ECC) won Anthropic’s hackathon and currently sits at 170K GitHub stars. One command installs a full configuration framework on top of Claude Code.
claude /plugin install everything-claude-code@everything-claude-codeThe 8 Modules at a Glance
| Module | Count | Loading | Token Impact |
|---|---|---|---|
| CLAUDE.md | 7 templates | Auto-loaded | High |
| Rules | 34 rules | Auto-loaded | High |
| Hooks | 20+ | Event-triggered | Medium (execution time) |
| MCP | 14 | Connected at startup | Low (startup latency) |
| Agents | 48 | On-demand | Zero (no cost until invoked) |
| Skills | 183 | On-demand | Zero (no cost until invoked) |
| Commands | 79 | User-triggered | Zero (no cost until invoked) |
| Memory | Project-scoped | Persistent across sessions | Low (only MEMORY.md index) |
The distinction that matters: the first four load automatically and eat tokens whether you use them or not. The last four cost nothing until you actually call them.
Module 1: CLAUDE.md
Your project’s employee handbook. Two layers: ~/.claude/CLAUDE.md (global) and ./CLAUDE.md (project-level, higher priority).
Seven templates ship out of the box: General Project, SaaS Next.js, Go Microservices, Django, Laravel, Rust, and Personal Global.
Keep the global layer under 200 lines. Write three things: who you are (role, industry, experience level), what the project is (tech stack, goals, constraints), and how you work (coding style, forbidden patterns).
Module 2: Rules (34 Rules)
Nine universal rules cover: coding style, testing, Git workflow, security, code review, agent usage, hooks, performance, and design patterns.
Then there are language-specific rulesets — 4 rules each across 11 ecosystems (TypeScript, Python, Go, Rust, Java, and more), totaling 44 language rules.
Optimization move: in your CLAUDE.md @import section, only pull in languages you actually use. This alone can cut 60% of wasted token overhead.
Module 3: Agents (48 Sub-Agents)
Each agent runs in its own isolated window with a separate 50K token context. Six categories:
- Code review (10): code-reviewer, typescript-reviewer, python-reviewer, security-reviewer, and others
- Build errors (7): build-error-resolver, test-failure-debugger, dependency-resolver, etc.
- Planning & design (3): planner, architect, tdd-guide
- Testing & refactoring (4): e2e-runner, refactor-cleaner, doc-updater, etc.
- Ops & coordination (3): chief-of-staff, loop-runner, config-tuner
- Specialized extensions (21): brand review, customer ops, cost auditing, UI/UX review, and more
You’ll use 5-10 daily. My active set: code-reviewer, security-reviewer, planner, architect, build-error-resolver, doc-updater, chief-of-staff.
Module 4: Skills (183 Capability Packs)
The difference between a Skill and an Agent: Skills execute in the main conversation window (lightweight, cheap). Agents spawn a separate window (heavier, isolated). If a Skill can handle it, don’t reach for an Agent.
Seven categories: core workflows (7), language patterns (14), framework patterns (20+), database & infrastructure (7), content operations (20+), security & quality (10+), and advanced specializations (100+).
Custom Skill template:
---
name: skill-name
description: One-line description for Claude to decide when to trigger it
---
## Trigger Conditions
When to invoke...
## Workflow
1. Step one
2. Step two
## Constraints
- Must do X
- Never do YModule 5: Hooks (The Soul of the System)
Hooks are real code, not prompts. The AI cannot modify them — that’s the fundamental difference from CLAUDE.md rules. Reliability is orders of magnitude higher than prompt-based instructions.
Six event types:
| Event | Fires when | Typical use |
|---|---|---|
| PreToolUse | Before tool execution | Block dangerous commands |
| PostToolUse | After tool execution | Auto-format, scan for console.log |
| SessionStart | Session initializes | Load context |
| Stop | Session ends | Save summaries, log token usage |
| UserPromptSubmit | User submits input | Inject reminders |
| PreCompact | Before context compaction | Preserve critical info |
Real example — fact-force: before any file write, forces the AI to declare 4 facts (who imports this file, which functions are affected, what data is read/written, and the user’s original instruction). Purpose: reduce hallucinations and prevent destructive changes.
Three strictness levels controlled via environment variable:
export ECC_HOOK_PROFILE=minimal # Only block dangerous operations
export ECC_HOOK_PROFILE=standard # Balanced (default)
export ECC_HOOK_PROFILE=strict # Confirm every changeModule 6: MCP (14 External Connections)
Code collaboration (GitHub, Supabase, Vercel), search and research (Exa, Context7, Brave), browser automation (Playwright, Browser Harness), local tools (Filesystem, SQLite), and AI enhancement (Sequential Thinking, Memory).
Each MCP takes 0.5-2 seconds to connect at startup. All 14 active means 7-28 seconds of startup latency. Remove MCPs you don’t use from ~/.claude/settings.json.
Modules 7 & 8: Commands + Memory
79 commands (/plan, /code-review, /tdd, /security-review, and so on). User-triggered, zero token overhead until invoked.
Memory comes in 4 types: user (personal preferences), feedback (corrections), project (decisions), and reference (pointers to external resources). Stored in ~/.claude/projects/<hash>/memory/, auto-loaded through a MEMORY.md index file.
6 Token Optimization Rules
| Target | Recommendation |
|---|---|
| CLAUDE.md | Keep global under 200 lines |
| Rules | Only import languages you use |
| Agents | Keep 5-10 in daily rotation |
| MCP | Delete unused connections |
| Hooks | Start with minimal profile |
| Skills/Commands | Don’t worry (zero cost until called) |
Think of ECC as a buffet, not a prix fixe menu. Take what you’ll eat. Leave the rest on the table.