Everything Claude Code (ECC)

Agent harness performance system by Affaan Mustafa (@affaanmustafa). MIT licensed. Claude Code plugin + npm package + manual install. Anthropic Hackathon win…

Everything Claude Code (ECC)

Agent harness performance system by Affaan Mustafa (@affaanmustafa). MIT licensed. Claude Code plugin + npm package + manual install. Anthropic Hackathon winner (Feb 2026).

• GitHub: affaan-m/everything-claude-code
• Plugin ID: ecc@ecc
• npm: ecc-universal

---

What It Is

A layered plugin for Claude Code (and 8 other harnesses) that adds: 60 agents, 232 skills, structured rules, hook-based automations, continuous learning, and security scanning. Framed explicitly as a “harness performance system” — the claim is that harness quality determines output quality as much as model quality does.

---

Install Paths

Plugin path (recommended for most users):

/plugin marketplace add https://github.com/affaan-m/everything-claude-code
/plugin install ecc@ecc
# Then manually copy rules (plugins can't distribute rules):
mkdir -p ~/.claude/rules/ecc
cp -r everything-claude-code/rules/common ~/.claude/rules/ecc/
cp -r everything-claude-code/rules/typescript ~/.claude/rules/ecc/  # pick your stack

Manual install (fallback):

./install.sh --profile full   # or: npx ecc-install --profile full

Do not stack methods — plugin + manual full install causes duplicates.

Selective (minimal):

./install.sh --profile minimal --target claude  # rules + agents only, no hooks

---

Cross-Harness Support Matrix

Harness	Mechanism	Hook parity	Agent parity
Claude Code	Native plugin	8 events	60 agents
Cursor	DRY adapter (transforms to CC format)	15 events	48 agents (prefixed ecc-*)
OpenCode	ecc-universal npm plugin	11 events (more than CC)	12 agents
Codex (app + CLI)	AGENTS.md + .codex/	None (instruction-based)	Shared via AGENTS.md
GitHub Copilot	.github/copilot-instructions.md + prompts	None	None
Zed	.zed/ adapter	Not available	Project-local
Antigravity	.agent/ adapter	Per Antigravity spec	Shared
JoyCode/CodeBuddy	Project-local selective install	Per CodeBuddy spec	Shared
Qwen CLI	Home-directory adapter	Per Qwen spec	Shared

DRY adapter pattern: Cursor’s adapter.js transforms Cursor’s stdin JSON to Claude Code’s format, allowing the same scripts/hooks/*.js to run on both harnesses without duplication. OpenCode’s plugin hook system has more events than Claude Code (11 vs 8 native types).

---

Component Summary

Component	Count	Surface
Agents	60	agents/*.md
Skills	232	skills/
Commands (maintained)	75	commands/
Rules (language packs)	7 dirs	rules/common, typescript, python, golang, swift, php, arkts
Hook scripts	20+	scripts/hooks/, hooks/hooks.json
Context injection modes	3	contexts/dev.md, review.md, research.md
CLAUDE.md examples	7	examples/

All counts self-reported from README (unverified independently).

---

Key Skills

Skill	Purpose
tdd-workflow	Red-green-improve cycle, 80%+ coverage
eval-harness	Eval-driven development, checkpoint/continuous modes
verification-loop	Build, test, lint, typecheck, security
strategic-compact	When/when-not to compact context
continuous-learning-v2	Instinct-based learning with confidence scoring
security-scan	Runs AgentShield from within Claude Code
search-first	Research-before-coding workflow
iterative-retrieval	Progressive context refinement for subagents
skill-create	Generate skills from git history

---

Hook Runtime Controls

export ECC_HOOK_PROFILE=minimal|standard|strict  # enforcement level
export ECC_DISABLED_HOOKS="pre:bash:tmux-reminder"  # comma-separated IDs
export ECC_SESSION_START_MAX_CHARS=4000  # cap context injection (default: 8000)
export ECC_SESSION_START_CONTEXT=off    # disable entirely for local/low-context

---

Token Optimization

Recommended ~/.claude/settings.json:

{
  "model": "sonnet",
  "env": {
    "MAX_THINKING_TOKENS": "10000",
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "50",
    "CLAUDE_CODE_SUBAGENT_MODEL": "haiku"
  }
}

Setting	Default	Recommended	Impact
model	opus	sonnet	~60% cost reduction (claimed)
MAX_THINKING_TOKENS	31,999	10,000	~70% reduction in hidden thinking cost (claimed)
CLAUDE_AUTOCOMPACT_PCT_OVERRIDE	95	50	Earlier compaction → better quality in long sessions
CLAUDE_CODE_SUBAGENT_MODEL	inherits	haiku	Subagents use cheaper model

All impact numbers claimed from ECC README; not independently verified.

MCP context cost: too many active MCP servers can reduce the effective 200K context window to ~70K (claimed, unverified). Recommended cap: under 10 MCPs, under 80 active tools.

Strategic compaction (via strategic-compact skill) — compact at: end of research/exploration phase, after completing a milestone, after a failed approach. Do NOT compact mid-implementation (lose variable names, file paths, partial state).

---

ECC 2.0 (Alpha)

Rust control plane in ecc2/ — builds locally, exposes dashboard, start, sessions, status, stop, resume, daemon. Not a general release; alpha quality. Tkinter desktop dashboard also available via npm run dashboard or python3 ecc_dashboard.py.

---

Related Wiki Pages

• AgentShield — ECC’s security scanning component
• Instinct Clustering (Homunculus Pattern) — continuous learning v2 pattern
• Claude Code → Cross-Platform Migration Matrix — cross-harness migration matrix
• Agent Harness — the broader harness concept
• Context Compression Strategies — strategic compaction (ECC’s strategic-compact skill)