ECC: Open-Source Agent Harness Fixes AI Coding Security & Memory (1.9k Stars in 24h)

The article opens with two pain points the author faced: a Check Point report (Feb 2025) showing Claude Code project files can be poisoned with malicious hooks that trigger remote code execution and steal API tokens (CVE-2025-59536), and OWASP publishing a Top 10 for MCP. The author found unverified shell scripts in their own .claude hooks directory. Simultaneously, every new Claude Code session starts with amnesia — previously learned rules, corrections, and hard-won fixes must be re-taught.

What ECC Is

ECC (Everything Claude Code, repo affaan-m/ECC, site ecc.tools) is not a model, completion tool, or UI. It is an agent harness — an operator system wrapping coding agents (Claude Code, Codex, Cursor, OpenCode). The README cites a paper claiming 98.4% of Claude Code is scaffolding; ECC targets that 98.4%. Launched from a hackathon in Jan 2025, it reached 214k+ stars and 32k+ forks in five months. Solo maintainer affaan-m, MIT license, weekly releases, covers seven harnesses. Monetization via GitHub App Pro tier ($19/seat/month); public repos run free.

Installation & Component Catalog

Two commands inside Claude Code:

/plugin marketplace add https://github.com/affaan-m/ECC
/plugin install ecc@ecc

After install you get a modular toolkit: 261 skills (YAML-frontmatter SKILL.md files covering TDD, security review, token optimization, MCP server authoring), 64 agents (dedicated workers for planning, review, build-fixing), 84 commands, plus hooks, rules, memory modules, and the AgentShield scanner. 175 contributors, ~19k monthly npm downloads. A consult command helps choose components:

npx ecc consult "security reviews" --target claude

Memory: Persist Everything, Optimize Context Window

ECC's memory philosophy: optimize the context window, persist everything else . Hooks automatically save session context and reload it on next start — no manual copy-paste. Continuous learning extracts corrections, failed commands, and encountered pitfalls from the session, turning them into instincts and new skills reusable on similar tasks. The workflow chain:

plan → test → implement → review → verify → remember → improve

. An agent finishes work, then a fresh context critiques its output; winning patterns become reusable processes. New sessions no longer start from zero.

Security: AgentShield with 102 Rules & Red/Blue Team Simulation

AgentShield scans CLAUDE.md, hook definitions, MCP server configs, and prompt files against 102 static rules, producing an A–F report. Runs entirely locally, no data upload. Rule library backed by 1,282 tests. Advanced mode spins up three parallel Claude Opus 4.6 instances: one red-team attacker, one blue-team defender, one independent auditor — hunting suspicious hooks, hidden injections, over-permissive permissions, leaked secrets. Threat model grounded in real incidents: AWS security advisory, Unit 42 (Mar 2025) observed indirect prompt injection against agents in the wild, Check Point RCE in Claude Code project files. The project's security guide states: assume malicious text will enter context, assume repo will be poisoned, assume memory will store wrong things; convenience layer must never run before isolation layer.

Cross-Tool Portability

ECC avoids vendor lock-in: AGENTS.md at repo root works across Claude Code, Cursor, Codex, OpenCode; SKILL.md format portable between Claude Code and Codex; Cursor lacks native hooks, so ECC provides a DRY adapter to reuse Claude Code hook scripts. v2.0 extends support to seven harnesses, adds Hermes Agent driver guide. README honestly notes non-Claude harnesses receive capability-reduced adapters, not full parity.

Caveats & Author's Verdict

ECC is not an app — no clickable UI. You adopt configurations and practices; without the accompanying guides, the files are just unfamiliar dotfiles. Deepest documentation lives in guides and the paid tier; the open repo is raw material. Supply-chain warnings on the install page: only trust github.com/affaan-m/ECC, npm ecc-universal, plugin ecc@ecc, and ecc.tools — third-party mirrors rejected. The author values the project not for star count but for one person distilling ten months of high-intensity daily use into copy-pasteable battle-tested patterns. They have personally audited the hooks.