ECC: one open harness for multiple AI agent environments

Background

Today's AI coding agent landscape is fragmented. Tools like Claude Code, Codex, OpenCode, and Cursor each have their own config files, plugin architectures, and follow-up patterns. In practice, that makes it hard to switch between environments: context management, security policies, and workflow orchestration end up scattered across handoffs. ECC is a MIT-licensed open-source platform that treats these overlapping environments as one unified operating layer.

What ECC solves

The biggest challenge ECC addresses is interoperability between harnesses. Most agent setups are tightly coupled to a specific tool, so changing the harness means rewriting rules and plugins. ECC reconstructs core components such as skills, instincts, memory, security, and research-first innovation in a harness-agnostic way. In other words, it is an agent harness performance optimization system — a framework designed to optimize the performance and operation of any agent runtime.

Here, performance means stability as much as speed. ECC's design philosophy assumes that if an agent can run code, it should be able to plan, test, review, debug, improve, and refine it autonomously through built-in pipelines. plan → test → implement → review → verify → remember → improve is not a loop rebuilt every time; it is the standard operating layer an installed agent relies on by default.

Structure and scope

The language breakdown is JavaScript 70.9%, Rust 20.9%, Python 4.6%, Shell 2.4%, TypeScript 1.1%, and Swift 0.1%. That ratio makes ECC's positioning clear. JS/TS handle plugin and environment integration, Rust covers performance and security workloads, and Python/Swift manage cross-platform adapters and integrations.

Supported tools include Claude Code as a first-class harness, Codex via a dedicated integration path, and compatible adapters for Cursor, OpenCode, Gemini, Zed, GitHub Copilot, Antigravity, and Qwen. Agent-specific directories such as .claude-plugin, .codex, .adal, .agents, and .gemini are included in the storage layout so the harness can be installed to match each agent's policy.

Package stats show 68 agents, 286 skills, and 94 commands in the repository, with a README that is actively maintained. Agents are specialized across planning, review, build operations, security, artifacts, and document work. Skills span TDD, research, security, documentation, prompt engineering, datasets, ML, and operations.

What changed in 2.0

ECC 2.0.0 marks stable graduation, meaning it is now a production-ready release. The biggest change is a session adapter layer that isolates the harness. ecc.session.v1 abstracts the session update process for Claude Code, Codex, OpenCode, and dmux under a common interface, greatly reducing the friction that arose from each tool having a different session storage approach.

MCP inventory support also improved. ecc.mcp.v1 audits MCP server settings across multiple harnesses in one pass. It detects scattered or conflicting configurations, identifies mismatches with the actual runtime, and reports drift. It also auto-resolves differences, so operators no longer have to compare each settings file manually.

Worktree lifecycle management is also present. When multiple agents work on the same workflow concurrently, the service predicts collisions and supports safe garbage collection. Actually tearing down and recreating a multi-agent workflow is the sharpest capability here.

Among orch-* orchestrator skills and rollout workflows, new optimization hooks have appeared: parallel execution optimization, benchmark optimization loops, accelerated data processing, and critical-path time reduction. Security hardening also continues to improve; curl-style flags are not persisted in argv, and gateguard has been adjusted to parse force and path checks via command-line arguments.

One notable change is the Node 21+ plugin engine cleanup. Changes that previously required an engine upgrade can now be made directly, though upgrades are still required for environments running older Node versions. As of late August 2026, version 2.2.0 was also released, continuing the multi-agent OS path with guided multi-harness installation plugins and unification initiatives.

Where it is useful

ECC is not a tool that forces a single agent on everyone; it is a layer that supports multiple agents simultaneously. For individuals working with two or more tools, switching paths becomes smoother. For teams using Claude Code, Codex, and OpenCode together, it is worthwhile to align settings, rules, and workflows under the ECC policy.

Depending on the installation environment, the number of plugins and commands may vary, so documentation should be treated as the ground truth. Because not all harnesses share the same feature set, checking platform support in the README before adopting a tool is recommended.

Installation is provided through a community channel. GitHub repositories, the npm packages ecc-universal and ecc-agentshield, a GitHub App, and the official site ecc.tools are community-supported entry points. Binary distributions and re-releases are available for convenience, and the project is actively maintained.

Active development is ongoing, with community documentation and a Discord community available. Before adopting, it is a good idea to confirm compatibility with your current environment.

In short, ECC is not a platform that forces a single agent harness; it is an approach that lets multiple agents share an execution layer. For teams running real multi-agent environments, checking MCP configuration, session management, and workflow rules through ECC is the first step worth trying.

References

• affaan-m/ECC GitHub repository
• Official site ecc.tools
• ECC 2.0.0 release announcement
• ECC 2.2.0 release notes