ECC Setup: Install the Claude Code Agent Harness in 10 Minutes

ECC is a free, MIT-licensed "operating system" for AI coding agents: one install drops 68 specialized agents, 286 skills, 94 slash commands, rules, hooks, and a security scanner into Claude Code, Codex, Cursor, and other harnesses, so your agent plans, tests, reviews, and remembers instead of freewheeling. It is trending on GitHub again today, adding more than 600 stars on top of roughly 245,000 total, and the whole setup is a single command that takes about ten minutes.

• ECC ships as an npm package (ecc-universal) and a native Claude Code plugin (ecc@ecc); the recommended path is one guided command: npx ecc-universal setup.
• Version 2.2.0 added guided setup for Claude Code, Codex, and Kimi Code in one wizard, plus adapters for Cursor, OpenCode, Gemini CLI, Zed, and Copilot.
• There is no API key, account, or subscription to configure: ECC layers on top of a Claude Code install that already works. An optional hosted GitHub App (ECC Pro) is the paid product.
• The main gotcha is stacking install methods: the plugin and the manual install.sh path install the same thing, and running both duplicates skills and fires hooks twice.

The exact steps, start to finish

1. Step 1. Check your prerequisites. ECC's guided setup needs Node.js 18 or newer, Git, and Claude Code 2.1 or newer on your PATH. Confirm all three before you start:

   # confirm Node, Git, and Claude Code are ready
   node --version
   git --version
   claude --version

2. Step 2. Run the guided setup. One command installs, updates, or moves the ecc@ecc plugin. It inventories your existing Claude Code scopes first, then asks which scope (user, project, or local) and which hook profile you want:

   npx ecc-universal setup

   If npm reports a version or cache error, check the registry before retrying:

   npm view ecc-universal version

3. Step 3. Add only the rule packs you use. Claude Code plugins cannot distribute rules, so rules are copied by hand. Clone the repo and copy the common pack plus one language pack (swap typescript for your stack):

   git clone https://github.com/affaan-m/ECC.git
   cd ECC
   mkdir -p ~/.claude/rules/ecc
   cp -R rules/common ~/.claude/rules/ecc/
   cp -R rules/typescript ~/.claude/rules/ecc/  # replace with your stack

4. Step 4. Verify the install. Run these from the same directory you used for setup. The first lists every ECC-managed target on the machine; the second reports missing or drifted files:

   npx ecc-universal list-installed
   npx ecc-universal doctor

5. Step 5. Use it for real. Open Claude Code inside any project, confirm the plugin loaded, then hand it a feature to plan. The planner agent returns an implementation blueprint you can edit before any code is written:

   /plugin list ecc@ecc
   /ecc:plan "Add authentication"

   That plan-first response is the payoff: from here, tdd-workflow enforces tests before code and /code-review reviews the result from a fresh context.

What is ECC and why is it trending?

ECC is a JavaScript project by developer Affaan Mustafa, created in January 2026 and released under the MIT license. Its pitch is blunt: an AI agent can already write code, but it forgets the plan, skips the tests, reviews its own work with the same context that wrote it, and relearns your conventions every session. ECC installs that missing process once. In the maintainer's words the loop is plan -> test -> implement -> review -> verify -> remember -> improve, and every piece of the package exists to make one of those steps mechanical rather than a reminder the model may ignore. The catalog is large: 68 agents (planner, architect, tdd-guide, code-reviewer, security-reviewer, and per-language reviewers), 286 skills, 94 command shims, opt-in rule packs, a hook runtime, an instinct-based learning system, and AgentShield, which scans your own agent configuration as an attack surface. Version 2.2.0 shipped guided setup for Claude Code, Codex, and Kimi Code plus a local operator control pane, and the repository sits near 245,000 stars with more than 37,000 forks. Today's spike tracks the 2.2 release and the wider shift toward agents that run whole workflows unsupervised.

RelatedSuperpowers Setup: Give Your AI Coding Agent a Real Workflow

How do you install ECC on Windows, macOS, and Linux?

The universal package is the same on every platform because it runs through Node. The command in Step 2 above is the whole story on macOS and Linux, and on Windows it runs identically from PowerShell or cmd. If you would rather not use npm at all, Claude Code's own plugin commands do the same job from inside a running session:

# inside Claude Code, the native plugin path
/plugin marketplace add https://github.com/affaan-m/ECC
/plugin install ecc@ecc

Pick one path per harness and stop. Both install the identical ecc@ecc plugin. The README also documents a low-context manual install for people who want rules, agents, and commands without the hook runtime. From a source checkout that is ./install.sh --profile minimal --target claude on macOS and Linux and, on Windows, its PowerShell twin:

# Windows, from the cloned ECC folder
.\install.ps1 --profile minimal --target claude

The catalog has seven install profiles. In our run, npx ecc-universal catalog profiles listed minimal (5 modules), core (6), developer (9), security (7), research (9), opencode (3), and full (26). Developer is the default most people want; minimal is the one to choose if you are watching context consumption.

How do you set it up for Codex, Cursor, and other editors?

Codex has a native plugin lifecycle now, so the install is two commands plus a check:

codex plugin marketplace add affaan-m/ECC
codex plugin add ecc@ecc
codex plugin list --json

Cursor, Gemini CLI, Zed, Antigravity, Qwen, Hermes, OpenClaw, Kimi, CodeBuddy, and JoyCode all use the same installer with a different target, run from the cloned repo. Cursor, for example, gets a project-local .cursor/ adapter:

./install.sh --profile minimal --target cursor

The README's support matrix is honest that these are capability-limited adapters: Claude Code gets the full experience today, Codex a supported native path, and everything else as much as that editor's plugin surface allows.

RelatedHallmark Setup: Kill AI Slop Design in Claude Code

How does ECC compare with rolling your own setup?

The honest counterargument is context. A hand-written CLAUDE.md is a few hundred tokens; ECC's rules are always loaded, and the full profile pulls in 26 modules. That is exactly why the maintainer tells you to start with rules/common plus one language pack and to run /context-budget when a session feels heavy.

What are the gotchas before you rely on it?

Four things worth knowing. First, do not stack install methods: the plugin path and the manual install.sh --profile full path install the same content, and layering them duplicates skills and makes hooks fire twice. If that already happened, remove the plugin, run node scripts/ecc.js uninstall --dry-run from the checkout, delete the rule folders you copied, and reinstall once. Second, hooks run shell commands and MCP servers hold credentials, so treat both as executable configuration; ECC's own GateGuard blocks destructive shell commands, and you can audit your whole agent surface any time with npx -y ecc-agentshield scan --path .. Third, doctor will nag you. On our test machine, an older manual 2.0.0 install reported 147 drifted files, and the guided setup refused to run at all until that overlap was removed, exactly the stacking the docs warn about; npx ecc-universal repair and the uninstall path exist for this. Fourth, hooks only load correctly on Claude Code 2.1 or newer, so check claude --version before blaming the plugin.

Our take

ECC's real product is discipline, not code. The individual pieces, a planner prompt, a TDD checklist, a reviewer that runs in a fresh context, are things any experienced Claude Code user has half-built in their own CLAUDE.md, and the value of ECC is that a single maintainer has been shipping the whole system weekly across seven harnesses while the rest of us kept rewriting the same three rules. That also defines its weakness: the catalog is now large enough that the "install everything" instinct actively hurts you. Install minimal or developer, copy one rule pack, and let the plan gate and the fresh-context review earn their keep for a week. If those two habits stick, the rest of the catalog is there when you need it. If they do not, you have lost ten minutes and learned something about how you actually work with an agent.

Original analysis by GenZTech. Tool documentation: affaan-m/ECC on GitHub.