Hey PH 👋 Three months ago an agent I was running went into a loop overnight and burned $200 in API credits. Last month a different one tried to run "git push --force" on production infrastructure. I couldn't find anything that blocked dangerous actions at the tool call level; not at the prompt level, not with a better system message. So I built Daedalab. It's a proxy that sits between your agents and the LLM providers. You change one environment variable. From that moment, every tool call gets classified in real time: 🟢 Safe → passes through 🔴 Critical (delete_file, git_push, send_email to real recipients) → blocked, Telegram notification fires, agent waits for your Approve or Deny The classification is deterministic, I read the function name from the tool call JSON directly. No LLM in the safety path. delete_file is always blocked. Every time. No exceptions, no hallucinations about whether something is dangerous. Works with Claude Code (API key), OpenClaw, LangChain, CrewAI, AutoGen; anything that exposes a configurable base_url. Happy to answer anything technical about how the proxy works. And yes, I know the irony of launching an agent safety tool the same week the terraform destroy incident went viral.

Daedalab

Your AI Agents, under control. Block before it's too late.

Your AI Agents, under control. Block before it's too late.