ECC：24万+ Star 的 Agent 工具箱，Claude Code / Codex 的工程操作系统
ecc-tools-agent-harness-286-skills-68-agents-claude-code
by Mycelium Protocol
GitHub：affaan-m/ECC ⭐ 243,000+ | MIT License | JavaScript
Website：https://ecc.tools
Discord：https://discord.gg/36yGMHGFbR
创建：2026-01-18 | 最近更新：2026-08-25
24 万 Star 说明了什么
ECC 在 GitHub 全球实时趋势榜常驻，7 个月从零长到 24 万+ Star，这个速度在非病毒式内容领域极少见。
它不是因为新奇的模型或者炫酷的界面被传播的。它被传播，是因为每天用 Claude Code 或 Codex 写代码的人发现：自己的 Agent 开始能按工程流程工作了，而不只是随机输出代码。
ECC 是什么
一句话：安装在你的 AI 编程助手之上的工程操作系统。
你的 Agent 本来能写代码。ECC 给它加上配套的工程系统和工具箱：
plan → test → implement → review → verify → remember → improve不是每次 prompt 里重新描述这个流程，而是安装一次，变成 Agent 的默认工作方式。
Optimize the context window. Persist everything else.
三层架构
ECC 不是单一仓库，是一个三层系统：
第一层：开源工具包（分发层）
- MIT 永久开源，这里是 ECC 的”前门”
- 286 个技能（TDD、安全、文档、前端、数据、ML、运维……）
- 68 个专项 Agent（规划、Review、构建修复、安全、架构、领域专家）
- 94 个命令快捷入口
- 跨平台适配：Claude Code、Codex、Cursor、OpenCode、Gemini、Zed、GitHub Copilot、Kimi Code……
第二层：AgentShield（安全保护层）
- 102 条安全规则，扫描每次 Agent 会话
- 检测范围：恶意 prompt、危险 hook、MCP 配置问题、权限泄露、secrets 暴露、Agent 文件篡改
- 开源扫描器，本地审计，可信透明
- GitHub App 自动化：PR 扫描 + 风险上下文 Review
第三层：ECC 2.0 控制平面（运营层）
- 本地优先的跨 Harness 控制平面
- 可观测性、编排、会话管理
- 跨 Claude Code / Codex / Cursor 的统一操作界面
核心能力一览
| 类别 | 数量 | 说明 |
|---|---|---|
| Agents | 68 | 规划、Review、构建修复、安全、架构、领域专家 |
| Skills | 286 | TDD、研究、安全、文档、前端、数据、ML、运维等 |
| Commands | 94 | 现有命令入口（ECC 正在向 skills-first 迁移） |
| Hooks & Memory | 运行时 | 强制执行、会话摘要、持续学习、本能、上下文控制 |
| Rules | 选装 | 按语言或项目选择的编码标准，每次加载 |
| AgentShield | 内置 | prompt、hook、MCP、权限、secrets、Agent 文件全扫描 |
安装
Claude Code（推荐路径）
/plugin marketplace add https://github.com/affaan-m/ECC
/plugin install ecc@ecc安装后 ECC 的 skills、agents、命令和 plugin-managed hooks 全部就位。选了这条路就停在这里，不要再手动安装。
npm（跨平台）
npm i -g ecc-universal支持 Codex、Cursor、OpenCode 等非 Claude Code 平台。
GitHub App
访问 https://github.com/apps/ecc-tools 安装，适合团队和 CI/CD 场景（私有仓库从 $19/seat/月起）。
为什么不只是一堆 CLAUDE.md
很多人试过自己写 CLAUDE.md 定义工程流程，效果参差不齐，原因不是 prompt 写得不好，而是：
- 每次新对话都要重建流程理解，没有持久化的”本能”
- 自定义 hook 很难写对，AgentShield 发现绝大多数自定义 hook 有安全漏洞
- 缺乏专项 Agent：规划 Agent 和实现 Agent 不一样，Review Agent 更不能用同一个上下文
ECC 解决的是这三个问题：通过 hooks + memory + instincts 把工程流程持久化，内置 AgentShield 扫描，拆分专项 Agent 角色。
商业模式
OSS 层永久 MIT 免费。ECC Pro（GitHub App）面向私有仓库团队，$19/seat/月，含自动 PR 扫描、策略包和企业级报告。赞助商包括 CodeRabbit、Greptile、Atlas Cloud、Moonshot AI（Kimi）等。
相关链接
- GitHub：https://github.com/affaan-m/ECC
- 官网：https://ecc.tools
- 技能目录：https://ecc.tools/skills
- 安全层：https://ecc.tools/security
- Discord：https://discord.gg/36yGMHGFbR
- GitHub App：https://github.com/apps/ecc-tools
Mycelium Protocol — 追踪 AI 系统的底层演化
关于 Mycelium
菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。
ECC: The 243k-Star Agent Harness OS for Claude Code, Codex, and Cursor
by Mycelium Protocol
GitHub: affaan-m/ECC ⭐ 243,000+ | MIT License | JavaScript
Website: https://ecc.tools
Discord: https://discord.gg/36yGMHGFbR
Created: 2026-01-18 | Updated: 2026-08-25
What 243k Stars Tells You
ECC has held a place on GitHub’s global trending list for months, growing from zero to 243k+ stars in seven months. That velocity is rare outside of viral meme repos.
It spread because people using Claude Code or Codex every day found something unexpected: their agent started working like a trained engineer, not just randomly emitting code.
What ECC Is
One sentence: an engineering operating system installed on top of your AI coding assistant.
Your agent can already write code. ECC gives it the coordination, process, and toolbox to do it like an engineer:
plan → test → implement → review → verify → remember → improveInstead of re-describing this process in every prompt, you install it once and it becomes the default.
Optimize the context window. Persist everything else.
Three Layers
ECC is not one repo. It’s a three-layer system:
Layer 1: Open-source toolkit (distribution)
- MIT-licensed forever — this is ECC’s front door
- 286 skills: TDD, research, security, docs, frontend, data, ML, operations, and more
- 68 specialized agents: planning, review, build repair, security, architecture, domain work
- 94 command shims as convenient entry points
- Cross-harness adapters: Claude Code, Codex, Cursor, OpenCode, Gemini, Zed, Copilot, Kimi Code
Layer 2: AgentShield (protection)
- 102 security rules scanning every agent session
- Scans: malicious prompts, dangerous hooks, MCP config issues, permission leakage, secret exposure, agent file tampering
- Open-source scanner for auditable trust; no automatic telemetry
- GitHub App automation: PR scanning and risky-context review
Layer 3: ECC 2.0 (control plane)
- Local-first cross-harness control plane
- Observability, orchestration, and session management above the underlying tools
- Unified operations surface across Claude Code, Codex, and Cursor
Core Inventory
| Category | Count | What it gives you |
|---|---|---|
| Agents | 68 | Planning, review, build repair, security, architecture, domain work |
| Skills | 286 | TDD, research, security, docs, frontend, data, ML, operations, more |
| Commands | 94 | Entry points while ECC moves to skills-first surface |
| Hooks & Memory | Runtime | Enforcement, session summaries, continuous learning, instincts, context controls |
| Rules | Selective | Always-loaded language or project standards you choose |
| AgentShield | Included | Prompt, hook, MCP, permission, secret, and agent-file scanning |
Install
Claude Code (recommended):
/plugin marketplace add https://github.com/affaan-m/ECC
/plugin install ecc@eccThis installs ECC’s skills, agents, commands, and plugin-managed hooks. Pick this path and stop — don’t also run a manual install.
npm (cross-platform):
npm i -g ecc-universalWorks with Codex, Cursor, OpenCode, and other harnesses.
GitHub App:
Install at https://github.com/apps/ecc-tools for team and CI/CD use. Private repos from $19/seat/month.
Why Not Just Write a Better CLAUDE.md
Many people have tried. Results are inconsistent for three reasons:
- Processes don’t persist across sessions — without hooks, instincts, and memory, every new conversation re-learns the process from scratch
- Custom hooks are hard to write securely — AgentShield finds security issues in the vast majority of hand-written hooks
- Planning and implementation agents are different roles — a single shared context can’t do both well
ECC addresses all three: persistent process via hooks + memory + instincts, built-in security scanning, and specialized agent roles that don’t share context.
Business Model
The OSS layer is MIT-licensed forever. ECC Pro (GitHub App) serves private-repo teams at $19/seat/month, with automated PR scanning, policy packs, and enterprise reporting. Sponsors include CodeRabbit, Greptile, Atlas Cloud, Moonshot AI (Kimi), and Itô Markets.
Links
- GitHub: https://github.com/affaan-m/ECC
- Website: https://ecc.tools
- Skills directory: https://ecc.tools/skills
- Security layer: https://ecc.tools/security
- Discord: https://discord.gg/36yGMHGFbR
- GitHub App: https://github.com/apps/ecc-tools
Mycelium Protocol — tracking the deep evolution of AI systems
© 2026 Mycelium Protocol. All rights reserved.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18，不接受商业合作、不代表任何企业或机构立场，也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息，而是给出有独立思考的深入分析，希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型，仅供学习交流与技术借鉴。它们大多仍处于早期阶段，有待进一步研究和验证，请勿直接用于工作或生产环境；如需采用，请先自行充分测试，并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理，不代表文中提及的任何公司、产品、模型的官方立场，未与其构成商业关联或合作关系。
- 科技行业信息更新极快，我们尽力保证内容准确、及时，但不对完整性、实时性做绝对保证，具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有，我们会尽量注明来源；如你认为存在版权疑问或侵权，请通过下方邮箱联系我们，收到通知后会尽快核实处理（更正、加注来源或删除）。
- 文章内容仅为技术科普与个人观点，不构成投资、法律或其他专业建议，据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询：hello@mushroom.cv
💬 评论与讨论
使用 GitHub 账号登录后发表评论