ECC — The Agent Harness Performance Optimization System(affaan-m/ECC)

One-line summary

ECC (Everything Claude Code) isn't just another AI coding plugin — it's a full performance-optimization system that "supercharges" your Claude Code, Cursor, and Codex agents. It packages scattered prompts, skills, memory, and security rules into reusable infrastructure, so your coding assistant arrives with experience and instincts instead of a blank slate every single session.

Background and origins

The story behind ECC is genuinely interesting. Its author, Affaan Mustafa, won first place at the Anthropic x Forum Ventures hackathon in September 2025 (held at Cerebral Valley in San Francisco) using this exact system, walking away with $15,000 in Anthropic credits. Over the following months he kept refining the build, distilling more than ten months of real, heavy daily use of AI coding assistants into a single open-source project, which he published on GitHub on January 18, 2026 under the MIT license.

The pain point is concrete. When people use Claude Code to write software, every new session means re-explaining project context, coding conventions, and architecture preferences all over again. A recent paper measured that Claude Code is about 98.4% scaffolding and only 1.6% model — meaning the quality of an agent depends almost entirely on the harness (the "tack" of system prompts, skills, memory, permissions, and security rules) wrapped around it, not the brain inside it. ECC's thesis is simple and sharp: stop rewriting the same prompts, and turn behaviors that already work into portable infrastructure. That landed hard. Less than five months after launch the repo crossed 230K stars, which tells you it hit a real nerve rather than a hype cycle.

Core features in detail

• Skills (skill library): 262 on-demand skills covering coding, research, security, content production, enterprise workflows, and agent orchestration. The key word is "on-demand" — a TypeScript project auto-loads the TS review skill, a Python test scenario triggers the TDD skill, so your context window never gets bloated with irrelevant stuff. Domain packs like clickhouse-io, backend-patterns, and frontend-patterns are matched dynamically to the detected tech stack.
• Instincts: a learning mechanism with a confidence score. The agent extracts repeated behaviors into "instincts" and reuses them next time without re-reasoning from scratch. You can run /instinct-status to see what's been accumulated, /instinct-export to move it to another machine, and /evolve to keep improving it. It's the practical version of recursive self-improvement applied to your everyday config.
• Memory: compressed cross-session memory that survives even when you close the session. Technical decisions, coding conventions, and architecture preferences are all retained, so the next conversation picks up where you left off. With ECC_AGENT_DATA_HOME you can give Cursor a separate data root so the two harnesses don't overwrite each other's session files on the same machine.
• AgentShield (security): 102 static-analysis rules and 1,282 tests scan for leaked secrets, misconfigurations, and injection risks in milliseconds. In --opus mode it even spins up three parallel agents — red team, blue team, and auditor — for adversarial security auditing. They actually take security seriously here, not as a checkbox.
• 64 specialized agents + 84 commands: from architect, code reviewer, and security auditor to build-error fixer and E2E tester, the entire dev-chain roster is covered. Dedicated workers like planner, tdd-guide, and refactor-cleaner handle specific jobs so the main agent can delegate instead of doing everything itself. It's like being handed a virtual team.

Architecture and implementation

Under the hood ECC is a Claude Code plugin that drives automation through a Hook event system — firing scripts at SessionStart, PreToolUse, PostToolUse, Stop, and other nodes. It uses a Manifest-driven install pipeline for "on-demand loading": a skill only gets injected into context when the matching tech stack is detected. Environment variables like ECC_SESSION_START_MAX_CHARS let you cap session-start context to prevent token overflow, and ECC_INSTINCT_CONFIDENCE_THRESHOLD controls how sure an instinct must be before it's injected. For multi-platform support, adapter.js translates Hook events from Cursor and friends into Claude Code's format, so one script set serves many tools. Installation also ships in minimal / core / full profiles, so you copy only the rules, agents, and skills you actually need.

Full comparison with competitors

To put it bluntly: claude-skills has more skills but leans generic, Superpowers leans engineering discipline, and ECC currently beats everyone on cross-platform consistency, production-grade security, and self-evolving memory. ECC even implements, as a community project, the exact harness-engineering principles Anthropic's own blog has been preaching — which is a big part of why it resonates with serious users.

Who it's for

• Heavy AI coding users: if you already live in Claude Code or Cursor and want to squeeze every drop of capability out of your agent, the payoff is biggest.
• Multi-stack teams: maintaining TS, Python, Go, Java, and Rust projects at once? Agents auto-match the right reviewer. Huge time-saver.
• Security/compliance work: finance and enterprise projects lean on AgentShield to prevent secret leaks and misconfiguration.
• Not for: casual ChatGPT askers. For light users this system is overweight — you'll feel the friction more than the benefit, and most of its power will sit unused.

Quick start guide

The simplest install is right inside Claude Code:

# 1. Add the ECC plugin marketplace source
/plugin marketplace add https://github.com/affaan-m/ECC

# 2. Install the plugin (auto-loads skills, commands, hooks)
/plugin install ecc@ecc

# 3. Manually copy rules (plugins don't auto-distribute the rules dir)
# copy the repo's rules/ into ~/.claude/rules/ecc/

# 4. Install the hook runtime via script (minimal profile is lightest)
./install.sh --profile minimal --target claude

Once installed, key commands are live: /loop-start launches the enhanced Agent Loop, /quality-gate runs quality gates, /harness-audit checks config health. To save context, set ECC_SESSION_START_MAX_CHARS=4000; use ECC_HOOK_PROFILE=standard to control hook strictness; and use ECC_DISABLED_HOOKS to turn off specific hooks you don't want.

Community and ecosystem

It started as a one-person project but now has 190+ contributors. The v2.0.0 release (June 2026) literally titled itself "The Agent Harness Operating System," with 14 version iterations along the way. The site ecc.tools ships a Dashboard GUI built as a Tkinter desktop app with theme and font customization. The ecosystem is still growing, but core quality still rests heavily on the author's 1,400+ commits — community contributions so far are mostly translations and supplementary skills rather than deep architectural changes.

Verdict

In my view ECC is the most thorough take yet on "AI coding scaffolding," and the cross-platform consistency plus self-evolving memory are genuine innovations. AgentShield's built-in security is also rare seriousness in this space. The downside is equally clear: the system is huge, anything beyond the minimal profile has a steep onboarding cost, and docs aren't friendly to beginners. Also, it's fundamentally designed for the Claude Code family — other harnesses get "best-effort alignment," not perfect parity. If you code heavily with AI, install it and you won't go back. If you're just kicking the tires, start with the minimal profile and don't load everything at once.