Show HN: Rampart – Open-source security for Claude and AI agents in YOLO mode

I've been running an AI agent 24/7 on my home lab through OpenClaw — k3s cluster management, shell commands, config edits, all unsupervised. I could see what the agent was running, but had no way to stop a bad command before it executed. So I built Rampart.

How it works: you write a YAML policy that says what's allowed, denied, or flagged. Rampart evaluates every tool call against that policy before it runs. Here's what a policy looks like:

  - rm -rf / → denied
  - sudo anything → logged for review
  - curl, wget → logged for review
  - git push, go build, normal dev commands → allowed
  - cat ~/.ssh/id_rsa → denied

Everything gets written to a hash-chained audit trail. You can watch it live with "rampart watch" or generate HTML reports with "rampart report".

Setup for Claude Code takes one command: "rampart setup claude-code". It installs hooks that intercept every Bash command, file read, and file write before execution. Blocked commands never run — Claude sees an error and moves on.

Setup for OpenClaw agents is also one command: "rampart setup openclaw". Works on Linux and macOS.

Also works as a shell wrapper for any agent ("rampart wrap"), an MCP protocol proxy ("rampart mcp"), or an HTTP API that agent platforms can consult before executing anything ("rampart serve").

Go, ~14K lines, Apache 2.0, zero runtime deps. Policy eval takes under 20 microseconds.

I'd love feedback on what policies you'd want out of the box and what integrations matter most.