Everything Claude Code hits 118K stars: is this the missing config layer for AI coding agents?

Every team I talk to that uses AI coding agents runs into the same problem: the tools ship with almost no configuration out of the box, so you end up rebuilding the same hooks, guardrails, and memory setup on every project. Everything Claude Code (ECC) is the most serious attempt I've seen at solving that at the community level.

It packages agents, skills, hooks, security scanning, and memory persistence into a single installable plugin that works across Claude Code, Codex, Cursor, and OpenCode. It just crossed 118,000 GitHub stars, which is a strong signal that this problem is more widespread than any one team's setup.

What Happened

Developer Affaan Mustafa released ECC as an open-source performance optimization layer for AI agent tools. The project repository now ships 30 specialized agents, 135 skills, 60 commands, and support for 12 programming language ecosystems.

What I'd note here is how far it's come since its origins as an Anthropic hackathon winner. It now has 136 contributors, 15,400 forks, 1,000+ commits, and translations in seven languages. That's not a side project anymore, that's a community that's actively invested in this becoming infrastructure.

The latest release, v1.9.0 (March 2026), added a selective install architecture, six new language-specific agents, and SQLite-backed session state. The selective install piece in particular is what makes this practical for real teams: you install only what your stack needs, not the entire thing.

Key Features

• 30 specialized subagents. Covers delegation across planner, architect, TDD guide, security reviewer, build-error resolver, and language-specific reviewers for TypeScript, Python, Go, Java, Kotlin, Rust, and C++. The depth here is what separates ECC from a basic prompt template.
• AgentShield security scanner. Runs 1,282 tests and 102 static analysis rules through a red-team/blue-team/auditor pipeline using three Claude Opus agents. This is the feature I'd prioritize evaluating first if your team is using MCP configs or CLAUDE.md in production.
• Instinct-based continuous learning. Extracts patterns from coding sessions, scores them by confidence, and clusters related instincts into reusable skills via the /evolve command. The agent gets better the more you use it, without manual prompt editing.
• Cross-tool parity. Supports Claude Code (8 hook types), Cursor (15 hook types), Codex (config.toml-based), and OpenCode (11 hook types). A shared adapter pattern lets Cursor reuse Claude Code's hook scripts directly.
• Hook runtime controls. Teams can tune strictness through environment variables without touching config files, which is useful for teams with different risk tolerances across environments.
• Token optimization defaults. Model routing defaults to Sonnet for 80% of tasks and to Opus for deep reasoning, with auto-compaction at 50% of the context rather than 95%. That alone can meaningfully reduce costs on long sessions.

Why It Matters

AI coding agents ship with minimal configuration out of the box. Most teams I've seen spend hours tuning prompts, writing hooks, and building guardrails, and then do it all over again on the next project. ECC turns that per-project effort into a shared, version-controlled system that travels with your team.

The security scanning piece is where I think ECC earns its keep most clearly. Misconfigurations in CLAUDE.md and MCP configs are a real and underappreciated risk, and having 1,282 automated tests catching them before they ship is the kind of guardrail most teams don't have time to build themselves.

Beyond the individual features, ECC reflects something I'm seeing more broadly: as AI coding agents move from single-file edits into multi-step workflows, the configuration layer around them matters as much as the model itself. ECC is one of the few projects treating that layer as a first-class engineering problem rather than an afterthought.

Example Use Case

A team building a Django REST API with Celery workers installs ECC with Python-specific rules using ./install.sh python. They get django-patterns, django-security, django-tdd, and django-verification skills loaded automatically, no configuration for rules they don't need.

Running /tdd invokes the TDD guide agent, which enforces a write-tests-first approach with 80%+ coverage targets. Before merging, /security-scan runs AgentShield against their Claude Code config, flagging exposed API keys in MCP server definitions or overly permissive hook permissions. The /learn command then extracts patterns from the session into instincts that carry into future work.

This is the workflow I'd point a Python team toward if they're spending more than a few hours a week on agent setup and prompt maintenance.

Competitive Context

ECC's primary target is Claude Code, but it explicitly supports the competition, and that's actually what makes it interesting from a tooling-strategy perspective.

Codex gets first-class support with AGENTS.md-based configuration. Cursor receives 15 hook event types through the shared adapter layer. OpenCode gets 31 commands and 6 native custom tools. For teams running Claude Code on the backend and Cursor on the frontend, ECC becomes a single configuration source that spans both, which removes a real coordination headache.

The gap ECC fills is standardization. Claude Code ships agents and hooks but leaves orchestration, security auditing, and cross-session memory to the user. ECC packages those missing pieces. That's not a criticism of Claude Code -- it's more that ECC is doing what mature ecosystems eventually produce: a community layer that fills in what the core tool intentionally left open.

My Take

If your team uses Claude Code, Codex, Cursor, or OpenCode in any serious capacity, ECC is worth an hour of evaluation time. The 30 agents and security scanners alone cover ground that most teams are either ignoring or reinventing manually. The selective install architecture means you're not taking on more than you need.

The repository is MIT-licensed and installs in under two minutes. At 118K stars and 136 contributors, this has enough community weight to be unlikely to disappear. I'd rather adopt it now than build the same thing myself six months from now.