I started digging into OpenClaw, then found this: https://clawhub.ai/steipete/trello. They are literally telling people that it's okay to expose credentials, just as long as it's in an env var.

If we’re building with AI, safety has to be a product principle, not an afterthought. AI makes coding more accessible, but it also strips away the guardrails you usually get from experienced mentors and hard-earned production lessons.

So we built Earl: an AI-safe CLI for LLM agents. Instead of giving agents raw shell + network access, Earl enforces template-driven calls, stores secrets in the OS keychain, blocks SSRF/private IP targets, supports egress allowlists, and sandboxes Bash/JS/SQL execution.

It also runs as an MCP server so agents can use those same controls natively.

Get started: https://github.com/brwse/earl?tab=readme-ov-file#quick-start