ops-guard-mcp

Stateful enforcement, VPS management, GitHub sync, and content pipeline — as an MCP server.

---

Most Claude Code users start with a CLAUDE.md full of rules: *"never run 2 agents at once," "always commit before pushing," "don't spawn on every input."* It works — until it doesn't. Claude reads the rules at session start, then promptly forgets them mid-task.

Phase 1 — MD rules. You write longer and longer instruction files. You add bold text. You add ⚠️ symbols. Claude still ignores them at the exact moment it matters.

Phase 2 — Python hooks. You discover that Claude Code has a hooks system. You write auto_commit.py, guard_agents.py, shell scripts that intercept tool calls. It's better — violations get blocked at the point of execution. But hooks are stateless. Each invocation spins up a fresh Python process. There's no memory of what happened 3 tool calls ago. And hooks can't query things — they can only block or allow.

Phase 3 — MCP server. Bernard was deep in hooks when he noticed something: hookify just runs Python. The hook system is a thin wrapper over subprocess calls. Why not go one level up and give Claude a proper API?

MCP gives you:

• Persistent state across the entire session (a running action log, a tweet queue)
• Query tools not just enforcement (check agent count *before* spawning, not just block after)
• Feedback loops (post_task_check scans your session and tells you what you missed)
• Structured data (git logs parsed into categories, env diffs as JSON)

The insight is that Claude follows tools better than rules. A tool call is an action in context. A rule is just text.

---

| # | To