On May 19th we covered Everything Claude Code when it was an interesting security tool with a scanner that audited your agent configuration. Today the repo has 197K stars, 27.5K forks, and is #1 on GitHub Trending. Something happened. And it wasn’t just marketing.
I want to explain what changed — because the distance between the ECC we covered before and today’s ECC isn’t incremental. It’s a rewrite of ambition.
The origin: an Anthropic hackathon in September 2025
Affaan Mustafa won an Anthropic hackathon with a simple but well-executed idea: AI agents are inconsistent not because models are bad, but because they lack structure. Without rules, an agent skips tests, modifies linter configs to silence errors instead of fixing them, makes commits with --no-verify, and loses context between sessions. The problem isn’t the model. It’s the engineering around the model.
ECC was born as the answer to that problem. A set of skills, hooks, rules, and specialized agents that give repeatable structure to any AI harness.
From 40K to 178K stars: what explains the jump
The growth wasn’t linear. There were two concrete accelerators:
First, the multi-harness expansion. ECC stopped being «the Claude Code plugin» to become a harness-agnostic standard. Today the same repo deploys configurations for Claude Code, Cursor, Codex, OpenCode, Gemini, Zed, and GitHub Copilot. A single AGENTS.md at the repo root is read by all the tools. That solved the problem any team using more than one harness faced — which is most teams in 2026.
Second, community maturity. 113 contributors, translations in 7 languages, and 768 accumulated commits turned ECC into something no dev wants to maintain alone. The .editorconfig of AI development, as the community describes it.
V1.9.0: what really changed
Version 1.9.0, released in March 2026 with 212 commits, has three changes that strike me as structurally relevant for teams:
1. Selective install architecture
Before, installing ECC was all or nothing. Now there’s a manifest-based pipeline — install-plan.js + install-apply.js — with a SQLite store that tracks what’s installed and supports incremental updates. A Python team doesn’t inherit TypeScript overhead. A Go shop installs exactly the Go rules, the Go agents, and nothing else.
For teams with multiple repos in different languages, this changes the adoption calculus. It’s no longer «we install ECC in the main repo and see»; it’s «we install what we need where we need it.»
2. 28 specialized agents
The jump from generic agents to 28 specialized subagents is the most underestimated change in the version. There are dedicated agents for planning, TDD enforcement, security review, build error resolution, and language-specific code review — TypeScript, Python, Go, Rust, Java, Kotlin, C++. These aren’t prompts with a different role. They’re agents with their own logic, their own rules, and their own hooks.
The one that interests me most as a practitioner: the TDD agent that blocks any flow that doesn’t write tests before code. It’s not a reminder. It’s a structural constraint.
3. Support for 12 language ecosystems
V1.9.0 added complete coverage for 12 ecosystems. What matters isn’t the number — it’s that each ecosystem has its own style rules, auto-formatting hooks, and adapted skills. The post:edit:format hook runs Black on Python, Prettier on TypeScript, gofmt on Go. Without manual configuration. Without «remembering to run the formatter.»
What didn’t change: the philosophy
What convinces me of ECC as a long-term bet is that the central philosophy didn’t yield to growth pressure. The project keeps treating agent inconsistency as an engineering problem, not a prompting one. Security hooks keep blocking --no-verify, detecting secrets in prompts (sk-, ghp_, AKIA), and protecting config files from being modified by the agent to silence errors.
197K stars didn’t make them add marketing features. They added infrastructure features.
My take: this is the new .editorconfig
There’s a pattern that repeats in tools that become standards: they start by solving one specific problem well, the community adopts them, and adoption turns them into the de facto interface everyone assumes is present. Prettier did it for formatting. ESLint did it for linting. ECC is doing the same for the agent behavior layer.
The question for teams in 2026 isn’t whether to adopt ECC. It’s when and with what scope. The selective install of v1.9.0 removed the last excuse not to start.
Resources
- Repository: GitHub - affaan-m/ECC: The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond. · GitHub
- Previous article on yoDEV (security and AgentShield): Everything Claude Code (ECC): El Scanner que Audita tu Setup de IA Antes de que lo Haga un Atacante
Has your team adopted ECC or some similar structure layer for your agents? What slowed down or accelerated that decision?