ECC

Product Introduction

1. Definition: ECC (ECC Tools) is an open-source, multi-layered agent harness performance optimization system designed for AI-powered coding assistants. It operates as a control plane and ecosystem of tools that sits atop individual coding harnesses like Claude Code, Codex, Opencode, and Cursor.
2. Core Value Proposition: ECC exists to transform AI coding agents from generic tools into team-aligned, continuously improving, and secure collaborators. Its primary value is in harness optimization, continuous learning from session history, and enterprise-grade security auditing for AI agent configurations, enabling teams to scale their use of AI coding assistants with consistency and governance.

Main Features

1. Open-Source Skill & Agent Registry: A central, MIT-licensed repository containing 261 pre-built skills, 64 specialized agents, and 84 commands. These are modular workflow patterns (e.g., /tdd-workflow, /security-review) that can be installed into compatible harnesses to enforce coding standards, testing protocols, and security practices. It works via configuration files (like CLAUDE.md or .cursorrules) that instruct the underlying AI model.
2. AgentShield Security Scanner: A configurable security auditor powered by a three-stage pipeline (Red Team, Blue Team, Auditor) utilizing models like Claude Opus. It scans agent configuration files (CLAUDE.md, .cursorrules, agents.json) for 102 security rules across 1282 tests, identifying vulnerabilities like unrestricted file system access, missing rate limiting, and prompt injection vectors. It outputs a scored audit report and can be run locally or integrated into CI/CD via the GitHub App.
3. Continuous Learning System (v2): An instinct-based learning loop that observes actual agent session transcripts (corrections, error-fix sequences) and extracts reusable, atomic behaviors. These "instincts" are assigned confidence scores that reinforce with use and automatically decay or are pruned if idle, preventing system bloat. This enables session-history learning beyond simple git diffs.
4. ECC 2.0 Operator Layer: A local-first control plane providing cross-harness observability and orchestration. It offers session and task visibility across multiple coding assistants (Claude Code, Cursor, etc.), token optimization, shared policy context, and a unified interface for managing the underlying tools, functioning as an agent harness operating system.
5. GitHub App Automation: A SaaS layer that analyzes a repository's git history to automatically generate custom skill files (SKILL.md, instincts.yaml) tailored to the team's established patterns. It creates a pull request for review, turning repo history into reusable defaults and providing a clear upgrade path from public repo evaluation to private repo coverage and enterprise governance.

Problems Solved

1. Pain Point: Inconsistent and insecure AI agent behavior. Without ECC, coding assistants lack team-specific patterns, have no memory across sessions, and operate with potentially vulnerable configurations, leading to security risks and inconsistent code quality.
2. Target Audience: Engineering Teams and DevOps adopting AI coding tools at scale; Security Engineers needing to audit and govern AI agent permissions; Open-Source Maintainers seeking to optimize contribution workflows; Enterprise Architects procuring and rolling out AI-assisted development platforms.
3. Use Cases: Enforcing Test-Driven Development (TDD) workflows automatically within Claude Code; conducting automated, security-focused code reviews on AI-generated pull requests; onboarding new team members with a standardized set of coding instincts and security rules; auditing and hardening the configuration of a fleet of Cursor or OpenCode instances across a large organization.

Unique Advantages

1. Differentiation: Unlike single-point solutions (e.g., a linter or a static prompt library), ECC provides a cohesive three-layer system: a free OSS distribution layer, an additive security/protection layer (AgentShield), and a paid control-plane layer (ECC 2.0). This contrasts with closed-platform agents by remaining harness-agnostic, working across Claude Code, Codex, Cursor, and OpenCode without vendor lock-in.
2. Key Innovation: The confidence-scored, auto-pruning continuous learning system that converts raw session transcripts into durable team knowledge. This moves beyond prompt engineering to create a self-improving, anti-bloating memory layer for AI coding workflows, a significant innovation in agentic workflow optimization.

Frequently Asked Questions (FAQ)

1. Is ECC Tools free to use? The core ECC open-source repository (skills, agents, AgentShield scanner) is completely free and MIT-licensed. The ECC Tools GitHub App offers a free tier for public repositories. Paid Pro ($19/seat) and Enterprise plans are for private repository coverage, advanced automation, and team governance features.
2. How does ECC's continuous learning work without compromising privacy? ECC's continuous learning system performs local ingest, keeping raw session transcripts on the user's machine. It uses rule-based, inspectable extraction to create anonymized instincts. Teams can opt into sharing aggregate usage insights, but code, content, and transcripts are never transmitted by default.
3. What is the difference between an ECC "skill" and an "agent"? A skill is a curated, reusable workflow pattern or command (e.g., /security-review) that modifies the behavior of the primary AI coding assistant. An agent is a specialized, often more autonomous sub-process (e.g., code-reviewer) designed for specific execution-heavy tasks like planning or review.
4. How do I install ECC for use with Cursor IDE? For Cursor, use the selective OSS install builder to generate a profile (core, developer, security, full). The installer detects Cursor and creates a .cursorrules file with the equivalent configurations. You can then audit this config using npx ecc-agentshield scan .cursorrules.
5. Can ECC be used with locally-hosted LLMs? Yes. The ECC OSS layer is model-agnostic and integrates via standard configuration files and the Model Context Protocol (MCP). Users have successfully deployed ECC skills and workflows with local LLM and MCP server setups, as noted in community discussions.