June 11, 2026
Serve and over-share
Show HN: A police department for your Claude Code agents
This AI watchdog sparked cop jokes, side-eyes, and a big "why not just lock it down?"
TLDR: This new tool keeps a detailed record of what an AI coding helper does and flags sketchy behavior later, but it doesn’t stop bad actions in real time. Commenters immediately split between mocking the cop branding and asking why anyone wouldn’t just prevent the behavior in the first place.
A new tool called agent-pd is pitching itself as a kind of internal affairs for your AI coding assistant: it watches what the assistant does, logs every move, and later flags suspicious behavior like snooping where it shouldn’t, trying forbidden commands, or wandering off task. The big catch? It doesn’t stop anything in the moment. It just records the action and blows the whistle afterward. And that detail is exactly where the comments lit up.
The mood in the community was a mix of "huh, neat" and "wait, why this metaphor?" One of the loudest reactions dunked on the whole "police department" branding, joking that cops aren’t exactly famous online for finding wrongdoing when they investigate themselves. That instantly turned the launch into a meme fest, with the product name getting roasted harder than the code. The vibe was less "wow, futuristic security" and more "bold choice, my guy".
Then came the practical crowd, asking the most Hacker News question possible: why use this at all if you can already restrict what the AI is allowed to do up front? In plain English, critics were basically saying, why install a security camera if you can just lock the door? That sparked the real divide: is after-the-fact evidence useful accountability, or just a fancy receipt for chaos? Either way, the comments made one thing clear: the tool may be about watching agents, but everyone is really watching the branding.
Key Points
- •agent-pd records every tool and permission event from Claude Code main agents and subagents, including Workflow-spawned subagents.
- •The tool is logging-only and does not block actions; blocking is left to Claude Code permission prompts or OS sandboxing.
- •Its hook writes normalized, hash-chained, per-session audit logs and captures events such as `PostToolUse`, `PermissionDenied`, `SubagentStart`, and `SubagentStop`.
- •The `pd` CLI analyzes logs with six deterministic detectors in Python at zero LLM token cost, while an optional `pd judge` adds an LLM-based pass.
- •The repository includes a reproducible demo showing log verification and an offense report with sample findings such as permission bypass detection.