Skip to content

ECC adds 72 agents and 328 skills to Claude Code

The MIT-licensed ECC agent harness packages planning, testing, review and memory into one install for Claude Code, Codex and other coding agents.

Draftship Editorial3 min read

A process layer for coding agents

On 2026-11-20 the repository affaan-m/ECC listed 307,367 stars and 45,241 forks, a gain of 3,563 stars over five days. The ECC agent harness is written mainly in JavaScript and released under the MIT license. Its pitch is narrow and practical: a coding agent can write code, but it does not plan, test, review and remember unless you keep asking it to, so ECC installs that routine once instead of rebuilding it in every prompt.

The README describes the intended loop as plan, test, implement, review, verify, remember, improve. ECC works best with Claude Code today, according to the project, has native Codex and Qoder plugin paths, and provides what it calls capability-limited adapters for Cursor, OpenCode, Gemini, Zed, GitHub Copilot, Antigravity, Qwen and other harnesses.

What the install actually contains

The README counts 72 agents, 328 skills and 99 legacy command shims, plus hooks, rules, memory, continuous learning and a security scanner called AgentShield. The agents are split across planning, review, build repair, security, architecture and domain work. The skills cover areas including test-driven development, research, security, docs, frontend, data, ML and operations. Rules are always-loaded standards you pick by language or project, and AgentShield scans prompts, hooks, MCP configuration, permissions, secrets and agent files.

Install paths and the compatibility gaps

The guided setup is npx [email protected] setup, which configures the Claude Code plugin and requires Node.js 18 or newer, Git, and Claude Code 2.1 or newer on the PATH. A second wizard, npx [email protected] install --guided, handles more than one harness in a single flow: it supports Claude Code, Codex and Kimi Code, preflights each selection before writing, and asks for a final confirmation.

Support is not equal across those three. Codex goes through its own marketplace and plugin lifecycle, with hook review and trust remaining Codex-owned. Kimi Code gets managed project files under ./.kimi-code, and ECC does not configure hooks, model or provider settings, or authentication for it. A further set of targets, including cursor, gemini, opencode, qwen, zed and antigravity, still uses documented ecc install --target paths until each adapter passes the project's collision, update, repair and uninstall checks.

The README is explicit that you should not stack install methods inside one harness, because a second install can duplicate skills, commands, hooks or configuration. It also separates three identifiers that are easy to confuse: the source repository is affaan-m/ECC, the Claude plugin is ecc@ecc, and the npm package is ecc-universal. Claude Code plugins cannot distribute rules, so rule packs are copied by hand, starting with a common set plus one language pack. The project warns that only its repository, the ecc-universal and ecc-agentshield npm packages, the GitHub App, the plugin slug and ecc.tools are official, and that mirrors are not reviewed.

Money and what stays free

The repository is MIT-licensed permanently, in the project's words. ECC Pro is a hosted GitHub App aimed at private repositories, with a free install and private repos from $19 per seat per month. The README says sponsors and Pro subscribers fund the work of a single maintainer shipping weekly across seven harnesses.

Limits and open questions

The README makes no benchmark claim. Nothing in it measures how much better an agent performs with ECC installed than without, so the argument is structural rather than demonstrated. The phrase capability-limited adapters is not broken down feature by feature, and readers are pointed to a support status matrix instead. Star growth signals attention, not quality, and the README itself notes that a pinned version is not a security audit.

What this means for people building apps with AI

The transferable idea is that the scaffolding around a model matters as much as the model. If you drive an app build through a coding agent, ECC's split between always-on rules, reusable skills and a separate review pass is a reasonable pattern to imitate by hand. If you build inside Draftship instead, the closest equivalent lever is how you write the instructions: see writing effective prompts. Draftship turns a plain-language description into a built and hosted app with a database and sign-in, and it is unrelated to ECC.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.