Everything Claude Code: The GitHub Repo Approaching 100K Stars That Turns Your AI Agent Into a Full Engineering Team
Some open-source projects collect GitHub stars on hype alone. Others earn them by solving a problem that tens of thousands of developers face every single day. Everything Claude Code falls squarely in the second camp. With 99,900 stars at the time of writing, 13,000 forks, and 113 contributors, this repository is on the verge of crossing the 100,000-star threshold, a milestone reached by only a handful of projects in GitHub's history.
The project began as one developer's obsession. Affaan Mustafa, a San Francisco-based developer, has been using Claude Code daily since its experimental rollout in February 2025. In September 2025, he won the Anthropic x Forum Ventures hackathon at Cerebral Valley with an agent optimization system he had been refining for months. Rather than keeping his work proprietary, he published the entire thing on GitHub in January 2026 under the MIT license. The repo took off immediately.
The value proposition is straightforward: turn Claude Code, Anthropic's coding agent, into a full engineering team. Where the standard tool gives you a single assistant that can read and write code, Everything Claude Code equips it with 28 specialized subagents, 119 reusable skills, 60 slash commands, 34 rules, over 20 automated hooks, and 14 MCP servers. The system also works with Cursor, Codex, and OpenCode across Windows, macOS, and Linux.
Version 1.9.0, released on March 21, 2026, brings the commit count to 768. This is no longer a side project. It is an ecosystem.
Why Developers Are Flocking to Everything Claude Code
The repo's success stems from a frustration shared across the developer community: Claude Code is powerful, but it forgets everything between sessions. It does not know your team's conventions. It does not remember the mistakes it made yesterday. It has no idea that your project uses hexagonal architecture or that your tests must follow a strict AAA pattern.
Everything Claude Code fills these gaps by layering a memory, learning, and automation system on top of the standard agent. Instead of repeating the same instructions at the start of every session, you install a profile that teaches Claude Code your standards, patterns, and expectations. The agent learns, adapts, and improves from one session to the next.
28 Agents, 119 Skills: How the Specialized Agent System Works
At its core, Everything Claude Code relies on a multi-agent architecture that mirrors the structure of a real development team. Each subagent has a specialty and activates when the context demands it.
Specialized agents by role
The system includes 28 subagents, each designed for a specific task. The code-reviewer analyzes quality, maintainability, and consistency. The security-reviewer checks compliance with OWASP Top 10 standards, detects exposed secrets, and enforces secure coding patterns. The tdd-guide enforces a test-driven development workflow where tests are written before implementation. Language-specific reviewers exist for TypeScript, JavaScript, Python, Go, Java, Rust, and C++. The planner breaks features down into phases, identifies dependencies, and establishes execution order.
The 119 skills and contextual loading
Skills are domain-specific knowledge files covering coding standards, backend and frontend patterns, security best practices, and testing conventions. The system uses a progressive disclosure mechanism: Claude first loads the names and descriptions of available skills, then loads into memory only those matching the current task context. This approach prevents saturating the context window with irrelevant information.
The NanoClaw v2 engine handles model routing and skill hot-loading. Each agent runs in an isolated environment via the Claude Agent SDK, ensuring clean context and memory management.
The 60 slash commands and hooks
Slash commands provide shortcuts for the most common workflows. The /tdd command launches a complete test-driven development cycle. /plan generates a structured implementation plan. /e2e creates end-to-end tests. /learn triggers a learning cycle. /harness-audit evaluates your harness configuration and produces a score.
The 20+ hooks automate behaviors before and after each tool call. They can auto-format code, run checks, and prevent recurring mistakes. The hooks.json file manages between 8 and 15 different events, including PreToolUse, PostToolUse, SessionStart, and SessionEnd.
Installation and Configuration: From Core Profile to Full Profile
One of the smartest design decisions in Everything Claude Code is its selective install system. Rather than forcing all 119 skills and 28 agents on every user, the project offers four installation profiles tailored to different needs.
The four installation profiles
The Core profile is the minimal foundation. It installs the fundamental skills and agents, enough to significantly improve Claude Code's behavior without excessive complexity. This is the recommended entry point for new users.
The Developer profile adds development-specific agents and skills: code reviewers, TDD guide, architecture patterns, and planning commands. It is the most popular profile among individual developers.
The Security profile integrates AgentShield and the full suite of security auditing tools. It targets teams working on sensitive projects or those that must comply with strict security standards.
The Full profile installs the entire system: all 28 agents, 119 skills, 60 commands, every hook, and every MCP server. It is designed for power users and teams that want to leverage every feature.
Installation methods
Installation can be done several ways. The fastest route is via npm with the npx ecc-install command, which launches an interactive assistant to choose your profile. The ecc-universal package (v1.0.0) installs 37 skills and 13 agents in a single command. The install.sh (Linux/macOS) and install.ps1 (Windows) scripts offer an alternative for those who prefer to avoid npm. Finally, the GitHub App enables direct repo-level integration.
Advanced memory configuration
Everything Claude Code provides memory compaction strategies that deserve attention. The system uses strategic-compact skills that limit thinking tokens (MAX_THINKING_TOKENS=10000) and adjust the auto-compaction percentage (CLAUDE_AUTOCOMPACT_PCT=50). The goal is to maintain a balance between context richness and context window efficiency, a critical concern for long sessions with AI agents.
Session state management relies on SQLite, enabling lightweight and portable persistence. Between sessions, the agent retains essential information about your project, conventions, and previously made decisions.
AgentShield: Security Built Into the Agent Layer
AI agent security is a topic most developers prefer to ignore until the day a misconfigured agent exposes secrets, writes to sensitive files, or executes destructive commands. AgentShield is Everything Claude Code's answer to this problem.
A security scanner with 102 static rules
AgentShield is an open-source security auditor powered by a red-team/blue-team/auditor pipeline built on Opus 4.6. It scans agent configuration files, including CLAUDE.md, .cursorrules, agents.json, and custom configs, looking for vulnerabilities, prompt injections, configuration drift, and guardrail gaps.
The scanner enforces 102 static rules and is backed by a test suite of 1,282 tests achieving 98% coverage. The command npx ecc-agentshield audit ./CLAUDE.md produces a detailed report with a security score out of 100, findings classified by severity (CRITICAL, WARNING, PASS), and actionable remediation recommendations. Reports export to JSON for easy integration into CI/CD pipelines.
Injection detection and guardrail verification
Beyond static analysis, AgentShield includes a prompt injection detection module that identifies attempts to manipulate agent behavior through malicious inputs. Guardrail verification ensures that user-defined restrictions are actually enforced by the agent at every session.
For enterprise teams, AgentShield can be deployed in PR review mode, where every pull request that modifies an agent configuration is automatically audited. This shift-left approach to AI agent security remains rare in the ecosystem, which partly explains the project's appeal.
Continuous Learning v2: When Your AI Agent Learns From Its Mistakes
The continuous learning system is arguably Everything Claude Code's most ambitious feature. Rather than updating model weights, which is obviously impossible for a third-party tool, the system creates and refines instincts: reusable behaviors that the agent acquires by observing your work sessions.
The instinct mechanism
Each instinct carries a confidence score ranging from 0.3 to 0.9. A newly created instinct starts with a low score. If it proves useful and relevant across sessions, its score gradually increases. If it produces poor results, its score drops until the instinct is automatically pruned.
This mechanism imitates, at a smaller scale, the learning process of a junior developer who gradually internalizes a team's conventions and best practices. The difference is that the agent can codify these learnings and share them across team members through the repo's files.
Learning commands
Four slash commands manage instincts. /instinct-status displays the current state of all instincts with their confidence scores. /instinct-export allows exporting instincts for sharing or backup. /instinct-evolve forces an evolution cycle that reevaluates and adjusts scores. /instinct-prune removes instincts whose scores have fallen below the relevance threshold.
The system's orchestration relies on a 5-layer observer loop that monitors each session, identifies recurring patterns, and transforms successful behaviors into persistent instincts. The harness audit scoring regularly evaluates overall system performance to identify improvement opportunities.
The Business Model: Free Open Source, Paid GitHub App
Everything Claude Code follows what has become a standard open-source playbook: the core is free, the added services are paid.
The open-source repository
The GitHub repository, under the MIT license, is and will remain free. You can install the entire system, including AgentShield, all 28 agents, and all 119 skills, without spending a cent. This free availability has fueled the project's explosive growth.
ECC Tools: the GitHub App
ECC Tools is the GitHub App that adds an automation layer on top of the open-source repository. It analyzes your repo's Git history and automatically generates skill files tailored to your team's patterns and conventions.
The Free plan covers public repo analysis with 10 analyses per month and 200 commits per run. It includes Core and Developer install profiles, issue-comment triggered PRs, and community support.
The Pro plan, at $19 per month per seat, unlocks private repo analysis, PR-triggered audits, AgentShield scanning, auto-triggers on large pushes and PRs, lifecycle tools (doctor, repair, uninstall), and priority support. It raises the limit to 50 analyses per month and 1,000 commits per run.
The Enterprise plan, available on request, offers unlimited analyses, SSO, audit logs, custom rule packs, dedicated onboarding, and a guaranteed SLA.
Genius or Over-Engineering? The Community Debate
Despite its nearly 100,000 stars, Everything Claude Code does not enjoy universal praise. A recurring debate runs through the developer community: is this system a brilliant advance or a textbook case of over-engineering?
The skeptics' argument
Critics point to the system's complexity. 28 agents, 119 skills, 60 commands, 34 rules, over 20 hooks: is all of this truly necessary? On Reddit, several experienced developers argue that Claude Code, used with solid commit messages, robust tests, and a well-structured CLAUDE.md file, does not need such an elaborate superstructure. The YAGNI principle, You Ain't Gonna Need It, comes up regularly.
Others note that many developers who adopt these complex systems do not truly understand what they are installing. The architecture looks robust, more polished than what they would have produced themselves, so they implement it without mastering all its moving parts. Days later, they find themselves debugging a system they do not understand.
The supporters' argument
Defenders counter that the complexity is justified by the use case. For a solo developer working on a simple project, the Core profile is more than sufficient. But for a team of ten working on a monorepo with security constraints, having specialized agents for code review, security, and TDD is not overkill. It is common sense.
The most compelling argument may be the modularity one. Thanks to the selective install profiles, nobody is forced to use the entire system. You can start with the Core profile and add components as your needs evolve. This progressive approach directly addresses the over-engineering criticism.
A question that extends beyond Everything Claude Code
The debate over AI workflow over-engineering extends well beyond this single project. It reflects a fundamental tension in the current ecosystem: AI agents are powerful enough to justify complex orchestration systems, but most developers have not yet built the intuitions needed to evaluate when that complexity is justified and when it becomes a hindrance.
What This Means for the Future of AI-Assisted Development
Everything Claude Code illustrates a broader trend: AI coding tools are no longer simple code-completion assistants. They are becoming orchestrated systems with specialized agents, learning mechanisms, and security layers. The fact that a community project can reach 100,000 stars in barely three months demonstrates the market's appetite for this kind of solution.
The next challenge for Affaan Mustafa and the community of 113 contributors will be turning this momentum into sustainability. The freemium model with ECC Tools is a start, but the real question is whether this type of harness will become an industry standard or remain a niche tool for power users.
At 99,900 stars, Everything Claude Code already has its answer. The community has voted. Whether the industry follows remains to be seen.
