ECC (affaan-m): 71 Agents, 311 Skills and What to Check Before You Install

ECC is the latest very large agent-harness collection to trend on GitHub, adding more than 1,000 stars in a day and standing near 280,000 in total. Published by affaan-m, it describes itself as an "agent harness performance optimization system" and ships 71 specialized subagents, 311 on-demand skills, hooks, rules, learned "instincts" and a memory vault. The current release is 2.2.3, dated October 1, and the license is MIT. Before you install anything with that many moving parts into the tool that has your shell access, read what it does and what its own README says is broken.

If skills are new to you, start with our complete guide to agent skills and the Claude Code commands reference.

TL;DR: questions answered

What is actually in the box

The README organizes ECC around a loop: plan, test, implement, review, verify, remember, improve. Its tagline is "Optimize the context window. Persist everything else." The parts:

• Agents (71): subagents for planning, review, build repair, security, architecture and domain work.
• Skills (311): on-demand workflows for TDD, security, research, documentation, frontend, data, ML and operations. Skills are the primary surface; the 97 commands are compatibility shims while the project migrates to skills.
• Hooks: runtime automation and enforcement, selectable by profile through ECC_HOOK_PROFILE.
• Instincts: patterns learned from sessions, each with a confidence score. Injection is capped by default at 6 instincts with minimum confidence 0.7.
• Memory Vault: ecc memory stores portable Markdown documents so context can move between harnesses. Project memories live in .ecc/memory/ and user memories in ~/.ecc/memory/. The README says memory is unreviewed context, not policy, which is the right way to treat it.
• Rules: opt-in standards with a common pack plus language packs. Plugins cannot distribute rules, so they are copied manually.
• AgentShield: scans prompts, hooks, MCP config, permissions and secrets.
• MCP: one default connector (chrome-devtools); others are opt-in, and Claude plugin installs do not auto-enable bundled MCP servers.

How to install with the smallest blast radius

The README's recommended path is a guided setup that needs Node.js 18 or newer:

npx ecc-universal@2.2.3 setup

The Claude plugin path additionally needs Git and Claude Code 2.1 or newer:

/plugin marketplace add https://github.com/affaan-m/ECC
/plugin install ecc@ecc

For Codex, the README lists codex plugin marketplace add affaan-m/ECC followed by codex plugin add ecc@ecc. A multi-harness wizard is available as npx ecc-universal@2.2.3 install --guided, and other harnesses use ./install.sh --profile minimal --target <name>. There are doctor, repair and uninstall commands with --dry-run options.

Our advice, not the README's:

1. Start with --profile minimal or --no-hooks, which skip the hook runtime, then add packs. With 311 skills, an unpruned install can crowd the context your agent needs for the real task.
2. Do not stack install methods. The README itself warns that doing so on one harness duplicates skills and hooks.
3. Pin the version, but also read the release source. The README says a version pin is not a security audit.
4. Try it in a throwaway repository or a container first, and diff your home directory afterward to see what was written.

Security: what the README says, and what you should add

ECC's own security notes are unusually direct. It tells you to install only from official channels (the GitHub repository, its two npm packages, the GitHub App, the plugin slug and ecc.tools) and says third-party re-uploads may contain malware. That matters because popular harness repos are natural targets for lookalike packages. Diagnostics are not uploaded automatically. The one data-sharing path to know about: the /astra-review command sends diffs to OpenAI through the local Codex CLI, and the README says invoking it is consent to that.

The structural risk is simple. Hooks run code on your machine at session events, skills shape what your agent does, and memory injects text into future prompts. A malicious or buggy entry in any of them is a prompt-injection or code-execution vector, which is the same class of issue we cover in the Capsem sandbox post and the ifixai agent audit guide. Run agents with the least permissions you can, and consider a runtime guard: AgentBeam, the agent security platform from the explainx.ai team, stops AI agents before they take dangerous actions.

What the README admits does not work yet

• Windows: core features are supported, but optional ones need Bash or Python. Open defects cover continuous learning and memory writes on native Windows (issues 2489 and 2626).
• macOS: the GAN shell path fails on system Bash 3.2 (issue 2674).
• Harness parity: Cursor, GitHub Copilot and others do not get full Claude feature parity. OpenCode ships only a subset of the catalog (issue 2617). Qoder enables only a reviewed SessionStart hook. Kimi hooks are not configured by the installer.
• External dependencies: multi-* commands need the separately installed ccg-workflow runtime. The ito-compute-cli package is unpublished and managed inference from the sponsor Itô Markets is described as not live yet.
• Counts: the README says 97 command shims in one place and 95 in a Qoder section, a small sign of fast-moving documentation.

The maintainer says they ship weekly across seven harnesses. That pace is both the appeal and the risk: frequent releases bring fixes, and also a steady stream of behavior changes in the hooks you run.

ECC versus other ways to extend your agent

Our top 25 Claude plugins and the Claude plugin developer portal cover the official route. For the broader field of harnesses, see open and closed source agent harnesses ranked and our loop engineering guide.

What people are asking

Is a bigger skill library better? Not by default. Every skill description competes for attention in the agent's context, and a skill that triggers at the wrong time can derail a task. Libraries this size are best treated as a catalog to pick from, not a default load. The README's own minimal profile and instinct cap (6 instincts, 0.7 minimum confidence) reflect that concern.

Why 97 commands if skills are the future? The commands are legacy shims kept for compatibility while the project migrates to skills. Expect them to shrink; avoid building muscle memory around them.

Can I use it with my team? The project offers sponsorship and a paid GitHub App, ECC Pro, with private repositories listed from $19 per seat per month. The core harness is MIT-licensed and free. If you roll it out to a team, check in a pinned configuration so everyone runs the same version, and review upgrades like any dependency.

Where does memory live, and who can read it? In plain Markdown files under .ecc/memory/ in the project and ~/.ecc/memory/ for the user. Because they are ordinary files, they can be committed by accident. Add the project folder to .gitignore unless you intend to share it, and never store secrets there.

How is it different from the official plugin directory? ECC is an independent project installed through its own marketplace entry, not Anthropic's vetted plugin directory. The project's site, ecc.tools, is one of the channels the README names as official, alongside the GitHub repository, the two npm packages, the GitHub App and the plugin slug.

A sensible adoption plan

1. Day 1: install minimal in a sandbox repo, run doctor, and read the hooks it registers.
2. Day 2 to 3: enable the TDD and review skills only. The README says the TDD workflow requires 80 percent or more coverage; check whether that fits your project.
3. Week 1: compare task success and token use with and without it on three repeatable tasks.
4. Week 2: decide whether to enable memory and instincts. Treat memory as untrusted context, as the README advises, and review what it writes.
5. Ongoing: subscribe to releases, re-run AgentShield after each upgrade, and remove anything you do not use.

Popularity is not evidence of fit. A 280,000-star repository tells you many people starred it; it does not tell you whether 311 skills improve your agent. Measure, prune, and keep the parts that earn their context.

Related reading

• What are agent skills? A complete guide
• Claude Code commands: complete reference
• Top 25 Claude plugins
• Agent harnesses ranked, open vs closed
• Capsem: VM sandbox for coding agents
• ifixai agent audit

Star counts, versions and issue numbers are from the GitHub README on October 15, 2026 and change often. Check the repository before installing.