GitHub - affaan-m/everything-claude-code: The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

Source
claude-codeai-agentsdeveloper-toolsprompt-engineeringcode-reviewtddmcp-serversai-workflow

Summary

Everything Claude Code (ECC) is a comprehensive, open-source performance optimization system for AI coding agents (Claude Code, Cursor, Codex, OpenCode) with 28 agents, 125 skills, 60 commands, hooks, rules, and MCP configs built over 10+ months of production use. It goes far beyond config files - providing continuous learning via instinct-based pattern extraction, cross-harness parity, security scanning (AgentShield), and token optimization strategies that can cut costs by 60%+.

Key Insight

  • Scale of the system: 28 specialized subagents (planner, architect, TDD guide, security reviewer, build-error resolvers per language), 125 skills across 12 language ecosystems, 60 slash commands. 997 internal tests passing.
  • Continuous learning is the standout feature: The instinct-based system (v2) auto-extracts patterns from sessions, scores them with confidence levels, and can cluster related instincts into reusable skills via /evolve. This is self-improving AI tooling.
  • Cross-harness parity matters: ECC works across Claude Code, Cursor (15 hook events vs Claude’s 8), Codex (app + CLI), OpenCode (11 events, 6 native custom tools), and Antigravity. AGENTS.md at root is the universal cross-tool file. DRY adapter pattern lets Cursor reuse Claude Code’s hook scripts.
  • Token optimization is critical and often missed: Each MCP tool description eats context window tokens - can reduce 200k to ~70k. Recommended: keep under 10 MCPs, under 80 tools active. Setting model: sonnet + MAX_THINKING_TOKENS: 10000 + CLAUDE_AUTOCOMPACT_PCT_OVERRIDE: 50 cuts costs ~60%.
  • Security scanning (AgentShield): Scans CLAUDE.md, settings.json, MCP configs, hooks, agent definitions across 5 categories. The --opus flag runs three Opus agents in red-team/blue-team/auditor pipeline for adversarial analysis, not just pattern matching.
  • Selective install architecture: Manifest-driven install pipeline with install-plan.js and install-apply.js for targeted component installation. State store tracks what is installed, enabling incremental updates.
  • Hook runtime controls: ECC_HOOK_PROFILE=minimal|standard|strict and ECC_DISABLED_HOOKS=... allow runtime gating without editing hook files.
  • Strategic compaction > auto-compaction: Compact after research/exploration before implementation, after milestones, after debugging - not mid-implementation where you lose variable names and partial state.