Who ECC is for#
Engineering teams standardizing AI coding conventions
Teams using Claude Code or Codex gain a centralized skills registry where conventions, TDD workflows, and security rules are published once and synced to every seat at session start. New developers pick up team patterns without manual onboarding documentation.
Skip if:
If your team has no shared coding conventions and no plans to build them, the skills layer adds configuration overhead without a matching benefit.
Security-focused developers auditing agent configurations
AgentShield scans CLAUDE.md files, hook configurations, MCP server settings, and prompt surfaces against 102 rules. Developers running AI agent tooling at scale can audit locally or integrate with the GitHub App for PR-level enforcement before conventions reach the broader team.
Skip if:
If your security requirements are already covered by existing tooling and you have no AI agent-specific exposure surface to audit, the scanning layer may not justify the setup cost.
Solo AI engineers building complex agent workflows
The 68 agents library covers specialized planning, build repair, research, and code review tasks. The continuous learning system observes sessions and promotes repeated workflows to named instincts with confidence scoring, so the harness improves over time without manual updates.
Skip if:
Developers who use a single harness casually may find the full install profile heavier than their workflow requires. The plugin install path is lighter than the full manual setup.
OSS maintainers evaluating repo automation
Public repos get the GitHub App for free. Commenting /ecc-tools analyze on any issue triggers an analysis of git history and generates a pull request with reusable skills and defaults. Teams can evaluate the full repo-native workflow before committing to a paid tier.
Skip if:
Repos with minimal git history or no established patterns will get sparse output from the analysis. The GitHub App's value scales with the richness of your existing commit history.
The problem it solves#
AI coding agents are powerful in isolation but stateless by default. Each session starts without knowledge of your team's conventions, preferred testing patterns, or past decisions. Developers configure their agents differently, so the same codebase gets inconsistent treatment depending on who is at the keyboard. Security checks, TDD workflows, and code review steps are ad hoc at best.
The challenge scales with team size. Without a shared skills layer or centralized convention management, onboarding a new developer means documenting how they should configure their agent setup separately from how the team actually works. Commercial orchestration platforms like Factory.ai and CrewAI address some of this but require adopting their specific runtime rather than extending the tools teams already use.
How it solves it#
Cross-harness compatibility
Works across Claude Code, Codex, Cursor, OpenCode, Gemini CLI, Zed, and more via a single install. Each harness gets a matching adapter or plugin so ECC's skills and agents load from the same source without duplication across setups.
286 skills and 68 agents library
Ships with 286 skills covering TDD, security review, research, documentation, frontend, data, ML, and operations, plus 68 specialized agents for planning, build repair, architecture, and code review. Teams publish skills centrally and every seat syncs on session start.
AgentShield security scanning
Scans every agent session against 102 security rules covering prompt injections, hook configurations, MCP server settings, permissions, secrets exposure, and agent files. Runs locally with no automatic data upload; integrates with the GitHub App for PR-level enforcement.
Continuous learning with instinct scoring
Observes sessions and extracts repeated corrections, error-to-fix sequences, and workflows into atomic instincts with confidence scoring. Idle instincts decay on a schedule; a reviewable prune plan shows what goes and why before anything is removed.
ECC 2.0 operator control plane
A local-first control plane for cross-harness session visibility, token optimization, operator review, and shared policy context. Session and task state are visible across multiple harnesses from a single surface, shipped as of ECC 2.0.0 stable.
GitHub App repo automation
Analyzes repository git history via a /ecc-tools analyze comment trigger and generates a pull request with reusable skills and defaults. Free for public repos; private repo coverage and deeper AgentShield-backed review checks start at $19 per seat per month.
Strengths and trade-offs#
Strengths
- MIT-licensed core with no per-seat costThe entire skills, agents, hooks, and commands library is MIT licensed. Teams can run, fork, and modify the OSS layer commercially without licensing fees. Unlike proprietary coding agent platforms, the core toolkit carries no subscription cost.
- Cross-harness portabilitySkills and agents install once and work across Claude Code, Codex, Cursor, and OpenCode. Teams that switch primary harnesses keep their conventions and institutional knowledge without rebuilding from scratch.
- Security governance included by defaultAgentShield scans every session against 102 rules covering prompts, hooks, MCP config, and secrets, with no separate subscription or setup. Commercial agent platforms typically charge separately for this level of session-level security governance.
- Large active OSS communityOver 242,000 GitHub stars and 36,000 forks as of August 2026, with enterprise sponsors including CodeRabbit and Greptile. The project ships weekly updates across seven harnesses, supported by a commercial GitHub App tier that funds ongoing development.
Trade-offs
- -Per-harness install paths are not interchangeableInstalling ECC across Claude Code, Codex, and Cursor requires separate steps for each harness. The README explicitly warns against stacking install methods, which can duplicate skills, commands, and hooks. Teams managing multiple harnesses face non-trivial setup coordination.
- -Claude Code is the primary harness; others are capability-limitedECC works best with Claude Code today. Cursor, Gemini CLI, Zed, and several other harnesses receive capability-limited adapters per the README, meaning some features are unavailable or degraded on non-Claude harnesses. Check the platform support matrix before assuming feature parity.
- -npm package lags the plugin install pathThe current npm release (ecc-universal 2.1.0) does not include guided setup commands. Teams using the npm path for Claude Code should switch to the native plugin commands until 2.2.0 ships. The README flags this gap explicitly.
ECC vs alternatives#
ECC vs GitHub Copilot
GitHub Copilot is a per-seat AI coding assistant focused on inline completions, chat, and multi-file edits inside editors. ECC is a harness layer that adds structure above the agent session, not a completion engine. The two serve different jobs and can coexist.
Teams running Claude Code, Codex, or Cursor can add ECC to gain session memory, skills libraries, and security scanning without changing their primary AI coding tool. Copilot's value is the completion itself; ECC's value is the convention, memory, and governance around it.
| Feature | ECC | GitHub Copilot |
|---|---|---|
| License | MIT (OSS layer) | Proprietary |
| Self-hosting | Yes | No |
| Session memory | Yes (continuous learning) | No |
| Security scanning | Yes (AgentShield, 102 rules) | Limited |
| Skills library | 286 skills, 68 agents | None |
| Cost | Free OSS; $19/seat GitHub App | $10-19/user/month |
ECC is the stronger choice for teams that need reproducible agent behavior, cross-harness portability, or session-level security governance. GitHub Copilot is the stronger choice for teams whose primary need is AI-powered inline completion inside VS Code or JetBrains without additional orchestration setup.
ECC vs Factory.ai
Factory.ai is a commercial AI software engineering service that runs managed agents on its own infrastructure. ECC extends the coding tools teams already use rather than routing work through a separate runtime.
Factory.ai handles infrastructure centrally and abstracts the underlying model. ECC is explicit about which harness it extends and what each install profile adds. Teams that want a fully managed engineering automation service without local infrastructure should evaluate Factory.ai. Teams that want to stay local, keep agent configurations auditable, and extend existing harnesses at no base cost should start with ECC's OSS layer.
| Feature | ECC | Factory.ai |
|---|---|---|
| License | MIT (OSS layer) | Proprietary |
| Deployment | Local harness extension | Cloud-managed |
| Agent runtime | Your existing harness | Factory's infrastructure |
| Base cost | Free (OSS) | Commercial, contact sales |
| Auditability | Full (local) | Limited (managed) |
Install and self-host#
Install ECC globally via npm for cross-harness setup across Claude Code, Codex, and Cursor.
```bash
npm install -g ecc-universal
```What it's built on#
- Languages
- JavaScriptPythonRustTypeScript
FAQ#
Is ECC free to use?
The OSS layer is MIT licensed and free. This covers the full skills library, 68 agents, hooks, and AgentShield scanner without any per-seat cost. The GitHub App is also free for public repos. Private repo coverage and deeper review features start at $19 per seat per month in the Pro tier. Enterprise pricing is available on request.
Which coding agents and editors does ECC support?
ECC has native support for Claude Code (via plugin commands) and Codex (via the repo-marketplace plugin). Cursor, OpenCode, Gemini CLI, Zed, and others are supported via install adapters. The README notes that non-Claude harnesses receive capability-limited adapters rather than full feature parity. Check the platform support matrix in the README before assuming a feature is available on your harness.
How does AgentShield work?
AgentShield scans agent configuration files, including CLAUDE.md, hook definitions, MCP server settings, and prompt surfaces, against 102 security rules. It runs locally with no automatic data upload. It integrates with the GitHub App for PR-level scanning when teams want automated enforcement on incoming configuration changes.
What is the difference between ECC skills and agents?
Skills are reusable instruction sets that define how the coding agent should behave in specific workflows, such as test-driven development, security review, or documentation generation. Agents are specialized subagents that handle a distinct task autonomously, such as planning a feature, repairing a build failure, or reviewing code quality. ECC ships 286 skills and 68 agents.
Does ECC require self-hosting?
The OSS layer runs locally with no remote dependency. You clone the repo, pick an install profile, and run the install command for your target harness. The GitHub App is a managed tier for teams who want repo-native automation and PR review integration without running infrastructure. The two paths can be combined: run the OSS layer locally and add the GitHub App for team-level governance.
Similar open-source tools#
codex
OpenAI's terminal coding agent, Apache-2.0 licensed
prime-agent
Automate coding workflows with a self-improving agent
VibePad
Control your AI coding assistant with a gamepad, open source
OpenCode
OpenCode is an open-source AI coding agent that assists developers in
go-modern-guidelines
Modern Go idioms for every AI coding agent, free
Superlearn
Interactive study boards for any topic, no new subscription needed