Affaan Mustafa’s Everything Claude Code (ECC) — an MIT-licensed agent-harness performance optimization system spanning 48 agents, 182 skills, 68 commands, 14 MCP servers, 34 rules, 997+ internal tests. Cross-harness by design: native parity testing across Claude Code (primary), Cursor IDE, Codex (macOS + CLI), OpenCode, Antigravity, Gemini CLI. Anthropic Hackathon winner.175,918 stars / 27,217 forks / 170+ contributors at fetch (2026-05-08), making it the highest-star Claude-Code-adjacent agent bundle on GitHub — created January 18 2026 (~4 months before fetch), with biweekly releases and rapid community response. Author Affaan Mustafa is a real founder (Itô / prediction-markets infra; ECC-Tools commercial entity; Twitter @affaanmustafa; blog affaanmustafa.com; 5,314 GitHub followers; 26 public repos; account from 2023-02-04). Distinct from agents (Claude-Code-marketplace-only) and agency-agents (multi-tool installer with personality-curated agents) by its explicit framing as a performance optimization system rather than a personality / capability bundle.
Key Takeaways
The “harness as performance system” framing is the unique pitch. ECC treats Claude Code, Cursor, Codex, OpenCode, Antigravity, Gemini CLI as agent harnesses with different capability ceilings, hook event counts, and rule formats — and ships configuration that maximizes performance per harness. Most other bundles in the wiki treat one harness as the deployment target. ECC’s 997 internal tests verify cross-harness behavior parity.
48 agents focused on code review, build resolution, and orchestration. Per-language reviewers (TypeScript / Python / Java / Go / C++ / Rust / Kotlin) and per-language build-error-resolvers (Java / Kotlin / Go / C++ / Rust / PyTorch). Plus development helpers (planner, architect, tdd-guide, e2e-runner, refactor-cleaner, doc-updater, database-reviewer) and orchestrators (loop-operator, harness-optimizer, chief-of-staff).
AgentShield is the security-scanning differentiator. Built at the Anthropic hackathon. 1,282 tests, 102 static rules. Scans CLAUDE.md, settings.json, MCP configs, hooks, agents, skills. Coverage: 14 secret-pattern categories, permissions, injection, MCP risks. CLI: npx ecc-agentshield scan [--fix] [--opus --stream]. No other bundle in the wiki ships dedicated security scanning for agent configs — closest analog is the Snyk ToxicSkills + Repello research which is a study, not a tool. ECC’s AgentShield is the practical defensive surface.
Memory persistence is “instinct-based” with v2 commands.
/instinct-status — view learned patterns with confidence scores
/instinct-import <file> — import patterns from others
/instinct-export — share your patterns
/evolve — cluster instincts into reusable skills
Pairs conceptually with Mahes’ Memory + Dreaming primitives in Claude Managed Agents — same shape (extract patterns, share across sessions, consolidate over time), different surface (ECC is local file-system-driven; Managed Agents is API-mediated).
Token-optimization recommendation: 60-70% cost reduction without quality loss.
Sonnet for 80%+ of tasks, thinking-tokens cap at 10k (vs 32k default), compact at 50% context (vs 95% auto), Opus only for deep architecture decisions. The 60-70% claim is ECC’s own; the effort/budget mechanic is Anthropic-canonical (see The Thinking Lever talk from Code with Claude 2026). ECC’s recommendations match Bleifer’s “use small/cheaper models on lower effort for the bulk of work” framing.
Hooks enforce discipline at the system level. 8+ hook event types: SessionStart (load prior context automatically), SessionEnd (save state, extract patterns), PreToolUse (block console.log / raw SQL), PostToolUse (auto-format / typecheck / lint), Stop (session summarization). Cursor and OpenCode get expanded surfaces (15+ and 20+ events respectively).
Critical setup gotchas (per README):
Don’t stack install methods (plugin OR manual installer, not both).
Rules cannot be distributed via plugins — copy manually from rules/ to ~/.claude/rules/ecc/.
Don’t add "hooks" to plugin.json — Claude Code v2.1+ auto-loads hooks/hooks.json.
Highest-volume Claude-Code-adjacent agent bundle on GitHub. Larger than agents (~34k★) and agency-agents (~95k★) by both star count and structural depth. Different framing: ECC is the performance system, agency-agents is the agent personalities, wshobson is the marketplace bundle. The three together cover the “curated agent infrastructure” landscape on Claude Code circa mid-2026.
Composes with oh-my-claudecode (OMC) — OMC orchestrates; ECC provides the rules/skills/agents OMC’s modes can recruit. Both ship multi-harness support but ECC’s harness-parity testing is more thorough.
Pairs with Superpowers as a meta-framework counterpart. Superpowers is the skill scaffolding language (define skills declaratively, layered with progressive disclosure); ECC is the curated bundle that ships those scaffolds populated.
AgentShield is the load-bearing security companion to Snyk ToxicSkills research. ToxicSkills found ~400 malicious skill-hub add-ons; AgentShield is the practical scanner you’d run before installing untrusted skills. Composable workflow: install ECC → run AgentShield → before adding any other skill bundle, scan it.
Memory + instinct system mirrors Mahes’ Memory + Dreaming at the harness-config layer. ECC’s /evolve command “cluster instincts into reusable skills” is structurally similar to Dreaming’s “consolidate patterns across sessions into a memory store” — same direction, different storage layer.
Token-optimization recommendations align with Bleifer’s The Thinking Lever. ECC’s “Sonnet at 10k thinking tokens, 50% compact threshold” is a concrete config that operationalizes Bleifer’s effort-and-budget framework.
170+ contributors and biweekly releases is a rare governance signal. Most curated bundles are single-author. ECC is a community-scale project with established release cadence — closer to Superpowers in shape (framework with community contributions) than to single-author bundles.
Setup — manual fallback:./install.sh --profile full (Mac/Linux) or .\install.ps1 --profile full (Windows).
Setup — non-Claude harnesses: ECC includes installers for Cursor, Codex, OpenCode, Antigravity, Gemini CLI — see harness matrix in the README.
Cost: Free (MIT). Token cost goes to your model provider; ECC’s recommended config targets a 60-70% reduction.
Integration notes:
Claude Code CLI v2.1.0+ required.
Rules can’t be distributed via plugins — must be copied manually from rules/.
Don’t add "hooks" to plugin.json — auto-loaded from hooks/hooks.json.
Don’t stack install methods — plugin OR manual, not both.
npx ecc-agentshield scan to security-audit your config / hooks / agents / skills before relying on them.
npm run dashboard (or python3 ecc_dashboard.py) opens the Tkinter GUI for component search.
npx ecc consult "<topic>" --target claude queries the catalog.
Selective install matters at this scale. 182 skills + 48 agents + 68 commands is too much to load wholesale. Use --profile flags or pick subsets.
Open Questions
175k-star adoption — organic vs trending. ~175k stars in ~4 months is extreme even for Anthropic-hackathon-winning projects. The 170+ contributor count and the codebase depth (75KB README, 997 tests, 14 MCP configs, biweekly releases) lean toward organic, but worth a last30days research check on community reactions.
AgentShield rule maintenance cadence. 102 static rules across 6 file types. Anthropic hooks/agents/skills surface evolves rapidly. How often do AgentShield rules update vs harness changes?
Cross-harness behavior parity in production. “997 tests” is a strong number but doesn’t say which behaviors. What’s the actual behavior delta between the same skill on Claude Code vs Cursor vs Codex vs OpenCode? Worth documenting.
Instinct/memory persistence durability. v1 was Stop-hook-extraction; v2 is /instinct-* commands. What’s the migration path for users with v1 instincts? Backwards-compatible? Required reformat?
Hook v2 reliability. README cites a “hook reliability overhaul with SessionStart fallback” in v1.8.0 — implies prior hooks were unreliable. What’s the current failure-mode taxonomy?
Selective install profiles. What’s in --profile full vs other profiles? Not enumerated in the README excerpt.
PolyForm-NC overlap with GitNexus. ECC is MIT (commercial-friendly); GitNexus is PolyForm-NC (commercial-restricted). For agencies pairing the two: ECC stack is fine commercially; GitNexus needs an akonlabs.com license. Worth noting in any “use both” recommendation.
Try It
Plugin install (5 minutes)./plugin marketplace add https://github.com/affaan-m/everything-claude-code then /plugin install everything-claude-code@everything-claude-code. Don’t forget to copy rules manually from rules/ to ~/.claude/rules/ecc/.
Apply the token-optimization config first. Set model: sonnet, MAX_THINKING_TOKENS: 10000, CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: 50 in your settings before doing anything else. Measure cost-per-task before / after for a representative session — ECC claims 60-70% reduction.
Run AgentShield once to audit your existing config.npx ecc-agentshield scan against your current ~/.claude/ directory. Look at the findings before adding anything new.
Try one verification skill.verification-loop is the highest-value cross-language skill — runs build/test/lint/typecheck/security gates as a single batched check. Use it once on a real PR.
For multi-harness teams, install ECC on whichever harness each team member uses. The harness-parity testing means the same skills behave consistently — useful when half the team is on Cursor and half on Claude Code.
For instinct extraction, run a real multi-day project, then /instinct-status to see what patterns ECC learned. /evolve to cluster them into reusable skills. Compare against Mahes’ Dreaming primitives from Code with Claude 2026 — different storage layer, similar idea.
For agency / WEO Marketly use, ECC is MIT — commercial use is fine without licensing friction. Combine with agency-agents for personality coverage and agents for marketplace breadth. Three layers, no overlap.
Read REPO-ASSESSMENT.md, SOUL.md, and WORKING-CONTEXT.md in the repo for the project’s own self-documentation. The 75KB README is the marketing side; the internal docs are the real shape.