AI coding tools like Claude Code, Cursor and Codex write code remarkably fast. The problem is that they work rather "on a whim": one day they plan carefully, the next they jump straight into code. They forget lessons from previous sessions, and you have to remind them every time: "remember to write tests", "remember the security review". ECC (github.com/affaan-m/ECC) was built to solve exactly this problem.
What is ECC?
ECC describes itself as an Agent Harness Operating System. It is an open-source framework (MIT license) developed by affaan-m that makes AI agents follow a disciplined engineering process instead of improvising differently every time.
ECC's philosophy fits in one sentence: "Optimize the context window. Persist everything else." Only what is truly needed goes into the context; knowledge, rules and workflows are persisted and reused.
Every task follows a closed loop:
plan → test → implement → review → verify → remember → improve
At the time of writing, the repository shows more than 270,000 stars and 40,000 forks, making it one of the most prominent AI agent projects today.
What's inside ECC?
| Component | Count | Role |
|---|---|---|
| Agents | 68 | Specialized subagents: planner, architect, code-reviewer, security-reviewer, language-specific reviewers for Go, Python, TypeScript, Rust… |
| Skills | 293 | Reusable workflows: TDD, security auditing, documentation, backend/frontend patterns, MLOps, database optimization… |
| Commands | 94 | Quick slash commands such as /plan, /code-review, /build-fix, /security-scan |
| Hooks | — | Event-triggered automation and enforcement |
| Rules | — | Universal and language-specific coding standards (opt-in) |
| AgentShield | — | Security scanning for prompts, permissions and leaked secrets |
Put simply: instead of one AI that "does everything", you get a whole team of virtual specialists. Each agent owns one stage and they collaborate through a clear process.
Notable features
1. Multi-tool support. ECC works best in Claude Code (as a plugin) and ships dedicated installers for Codex, Cursor, OpenCode, Gemini, Zed, GitHub Copilot, Kimi Code, Qwen, Antigravity and more. Switch tools and your workflow stays the same.
2. Research first, code later. Agents are guided to look up documentation and existing solutions before implementing, reducing "reinventing the wheel" and the use of outdated APIs.
3. Continuous learning. ECC extracts patterns from previous work and turns them into reusable "instincts". The more you use it, the better the agents understand your project.
4. Built-in security. AgentShield scans prompts, permissions and secrets — critical as AI agents are granted more and more access to source code and infrastructure.
5. Self-hosted model support. ECC works with custom API endpoints, a good fit for enterprises that must keep data in-house.
Quick installation
Recommended (Claude Code): npx ecc-universal@2.2.2 setup
Via the Claude Code plugin system: /plugin install ecc@ecc
Manual install (choose the minimal, core or full profile): ./install.sh
The latest release, v2.2.2, adds a guided setup for Claude Code, Codex and Kimi Code, with manifest-driven installation, health checks and safe uninstall.
Who should use ECC?
- Individual developers who want AI-written code that comes with tests and reviews, with less cleanup afterwards.
- Development teams that need to standardize how everyone uses AI, so results follow the same process and standards no matter who runs it.
- Enterprises that care about security, control over agent permissions and the ability to run on internal models.
Conclusion
ECC highlights a clear trend: the race is no longer just about which AI writes better code, but which AI works with a better process. With planning, tests, reviews and long-term memory, AI coding agents start to look more like real engineers than autocomplete tools.
If you use Claude Code or Cursor every day, ECC is well worth a try. Start with the minimal profile, then gradually enable the agents and skills that fit your project.
👉 Repo: https://github.com/affaan-m/ECC
