A 249-skill harness for agents that can't remember their job

What it does

ECC is an operator system for AI coding agents. It bundles hundreds of pre-built skills, session hooks, security rules, and MCP configurations for Claude Code, Cursor, Codex, OpenCode, and similar harnesses, aiming to make agentic workflows production-ready rather than experimental. The repo also includes a Tkinter dashboard, a Rust control-plane prototype, and a GitHub App for private repository audits.

The interesting bit

Instead of treating each IDE agent as a standalone tool, ECC frames them as a single harness that needs shared memory, token optimization, and continuous learning. The documentation is literally hosted as X threads, which feels either brilliantly concise or mildly unhinged.

Key highlights

• Cross-harness support for Claude Code, Cursor, Codex, OpenCode, Gemini, Zed, and Copilot with tightened behavior parity across all of them.
• 249 skills, 63 agents, and 79 legacy command shims covering framework patterns, media tooling, and prediction-market research.
• SQLite-backed memory persistence and session hooks that automatically save and reload context across restarts.
• AgentShield security scanning, sandboxing, and cost-audit controls baked into the workflow layer.
• Manifest-driven selective install architecture so you can ingest only the agents and skills you actually need.

Caveats

• The main guides explaining the system are X/Twitter threads, not in-repo documentation, which makes deep research awkward.
• The ECC 2.0 Rust control plane is explicitly alpha-quality and “not yet a general release.”
• The project is maintained by a single developer shipping weekly across seven harnesses, so bus factor is a real concern.

Verdict

Worth exploring if you live inside AI agent harnesses daily and need opinionated guardrails, memory, and skill reuse across multiple tools. Skip it if you want a lightweight drop-in config or a mature framework with a broad maintainer pool.