ECC (Everything Claude Code)

What It Does

ECC ("Everything Claude Code", later rebranded "the agent harness operating system") is a single repository that bundles 278 agent skills, 67 sub-agents, 94 slash commands, 122 rule files, and a manifest-tracked installer that fans the bundle out across 14 different agent harnesses. It targets developers who live inside coding agents (Claude Code, Codex, OpenCode, Cursor, Gemini, Hermes, OpenClaw, and others) and want a prebuilt operator layer covering TDD discipline, code review, security scanning, language-specific patterns for a dozen ecosystems (Python, Rust, Go, Kotlin, Swift, Laravel, Django, Spring, Quarkus, Angular, React, Vue), plus research and orchestration workflows. At v2.0.0 (released 2026-06-10) it is the second-most-starred agent skill repository on GitHub at 232,374 stars, behind only obra/superpowers.

The Good

The test suite is real and it passes. npm test runs a chain of structural validators (unicode safety, agent/command/rule/skill/hook/manifest validation, no-personal-paths, catalog, command-registry) followed by tests/run-all.js, and it returns 3,146 passed of 3,146 with exit 0. That is not typical for a skill pack of this size, where "tested" usually means the README says so. Here the installer, uninstaller, plan canvas, control pane, work-item claims, the NanoClaw REPL, the Codex sync script, and the Trae install scripts each have their own unit suites, and they are green.

The installer is manifest-tracked, reversible, and Hermes-first. node scripts/install-apply.js --dry-run --json --target hermes --profile minimal resolves a real adapter (hermes-home, kind home) targeting ~/.hermes, then plans 454 file copies across five modules. Running it for real into an isolated HOME wrote 455 files with zero errors and a 186 KB ecc-install-state.json manifest. ecc doctor --target hermes reports Status: OK, Issues: none, and uninstall --dry-run plans exactly 455 removals, so the install is fully reversible rather than a one-way copy. Seven profiles (minimal, opencode, core, developer, security, research, full) and 32 modules let you install a slice instead of the whole thing.

Cross-harness coverage is best in class. The installer lists 14 targets (claude, claude-project, cursor, antigravity, codex, gemini, opencode, codebuddy, joycode, qwen, zed, hermes, openclaw, kimi), the repo ships harness config dirs for all of them, plus a dedicated docs/HERMES-SETUP.md, a Claude Code marketplace.json, and an agent.yaml gitagent export. For Hermes users specifically this is first-class: a real adapter, a setup guide, and a profile system, not a copy-paste of the Claude config.

Every agent carries a Prompt Defense Baseline. All 67 agent markdown files open with a shared anti-injection section (reject role overrides, treat fetched and URL content as untrusted, treat homoglyphs and zero-width characters as suspicious, detect repeated abuse). Combined with a SECURITY.md policy, a supply-chain IOC scanner (scripts/ci/scan-supply-chain-iocs.js), and a top-of-README malware warning about third-party mirrors, the security posture is more deliberate than most skill repos.

The Bad

Breadth comes at the cost of depth, and progressive disclosure is mostly absent. Only 7 of 278 skills use a references/ directory (59 reference files total). The median skill is 190 lines of single-file markdown, which is fine for a focused contract but means the marquee skills (TDD, security review, eval harness) are not progressively disclosed the way a mature skill like graphify or the anthropics/skills pack is. Some skills are intentionally thin: nanoclaw-repl is 34 lines. That one is honestly scoped (operate a zero-dependency REPL), but the overall distribution is wide rather than deep.

The documentation has stale references to the old repo name. Roughly 20 files still point at github.com/affaan-m/everything-claude-code.git, including the flagship configure-ecc skill (its Step 0 clones that URL). The old name does redirect to ECC, so the clone still works, but a first-time reader following the bootstrap instructions hits a name that no longer matches the project. The README also hardcodes "211.9K+ stars" and "32.5K+ forks" in badge text, understating the actual 232,374 stars and 35,421 forks. Neither is fatal, but both are exactly the kind of drift that a 3,146-test suite does not catch because the tests do not cover marketing copy.

Commercial messaging is interleaved with the tool surface. The installer help text (install-apply.js --help) ends with an "Ito compute sponsor" promotion block, and the README front-loads an ECC Pro upsell (private repos, GitHub App, $19 per seat per month) alongside sponsor calls. None of it is hidden, and the core bundle stays MIT and free, but a reader who expected a pure open-source tool gets a product funnel baked into the CLI help.

Smoke Test Results

Testing ran on the host (macOS 26.5.1, aarch64, Node 24.13.1). The repo was cloned shallow (79 MB, 3,332 files) and dependencies installed via npm install (210 packages, 5 seconds) before the validators and test suites ran.

Run A. Fresh sandbox, no deps preinstalled

$ git clone --depth 1 https://github.com/affaan-m/ECC.git
✅ Cloned, 3,332 files (excl .git), 79 MB

$ cat VERSION
✅ 2.0.0

$ npm install --no-audit --no-fund --loglevel=error
✅ added 210 packages in 5s, exit 0

Pass rate: 3 of 3. The clone resolves cleanly and the dependency install (3 runtime deps: @iarna/toml, ajv, sql.js) has no friction.

Full fresh log →

Run B. Sandbox with deps preinstalled

$ npm test
✅ validators (unicode/agents/commands/rules/skills/hooks/manifests/no-personal-paths/catalog/command-registry) all pass
✅ tests/run-all.js: install-apply, uninstall, plan-canvas, control-pane, work-items, claw, sync-ecc-to-codex, trae-install suites green
✅ Final: Total 3146, Passed 3146, Failed 0, exit 0

$ node scripts/install-apply.js --dry-run --json --target hermes --profile minimal
✅ Resolves hermes-home adapter, 454 planned copy-file ops, 0 warnings

$ HOME=/tmp/ecc_test_home node scripts/install-apply.js --json --target hermes --profile minimal
✅ 454 ops applied, errors null, warnings []. 455 files written under ~/.hermes

$ HOME=/tmp/ecc_test_home node scripts/doctor.js --target hermes
✅ Status: OK, Issues: none

$ HOME=/tmp/ecc_test_home node scripts/uninstall.js --target hermes --dry-run
✅ Planned removals: 455 (matches written files), 0 errors

Pass rate: 5 of 5. This is the canonical install-correctness verdict. The full test suite is green and the Hermes install, doctor, and uninstall cycle is verified end to end against real file writes.

Full with-deps log →

Run C. Functional verification (does it do what it claims?)

$ find skills -name SKILL.md | wc -l
✅ 278 (README claim: 278 skills)

$ find agents -type f | wc -l
✅ 67 (README claim: 67 agents)

$ grep -c "Prompt Defense Baseline" agents/*.md | grep -v ':0' | wc -l
✅ 67 of 67 agents carry the anti-injection baseline

$ node scripts/install-apply.js --help | grep -A3 "^Targets:"
✅ 14 targets listed, hermes present; docs/HERMES-SETUP.md exists (5033 bytes)

$ grep -rl "everything-claude-code" --include="*.md" --include="*.js" . | wc -l
❌ 20 files reference the old repo name (stale; clone still redirects)

$ grep -c "211.9K" README.md
❌ README badge text understates stars (says 211.9K+, actual 232,374)

Functional pass rate: 4 of 6. The structural and install claims all check out. The two failures are documentation drift (stale URLs, stale star badge), not functional breakage.

Full functional log →

What the runs tell you

The runs say the engineering underneath ECC is solid: a green 3,146-test suite, a real reversible installer, and a verified Hermes path. The weaknesses are in the content layer and the marketing layer, not the plumbing. The skills are correct in format but shallow in progressive disclosure, and the docs trail the project's own rename and growth.

Setup Walkthrough

1. Clone the repo: git clone --depth 1 https://github.com/affaan-m/ECC.git && cd ECC.
2. Install Node dependencies once: npm install (Node 18+ required, 210 packages, ~5 seconds).
3. Preview what will land before touching your home directory: node scripts/install-apply.js --dry-run --target hermes --profile minimal.
4. Install for real: bash install.sh --target hermes --profile minimal (writes under ~/.hermes).
5. Check health: node scripts/doctor.js --target hermes (reports OK with no issues on a clean install).
6. To remove everything later: node scripts/uninstall.js --target hermes reverses the install using the state manifest.

Post-install gotcha: the .hermes/ directory shipped in the repo is a README only. The real Hermes skills, rules, and agents are materialized by the installer under ~/.hermes/.agents/skills/, ~/.hermes/rules/, and ~/.hermes/AGENTS.md, so do not expect to copy .hermes/ directly.

Alternatives

1. anthropics/skills - the official Anthropic reference pack. Fewer skills, deeper per skill, no installer funnel. Prefer it when you want canonical depth over breadth.
2. obra/superpowers - the #1-starred agent skill repo, a methodology-first framework. Prefer it when you want a routed workflow philosophy rather than a large flat skill catalog.
3. mohitagw15856/pm-claude-skills - 466 single-author PM skills with a zero-dependency MCP server. Prefer it when the use case is product management and project delivery rather than full-stack engineering.
4. mattpocock/skills - a routed idea-to-ship workflow with a router skill and a skill-authoring vocabulary. Prefer it for engineering workflow routing with a smaller, opinionated surface.