Agent Guard and loop watch

Irreversible calls stop before they run.

Before Claude Code or Codex runs a tool call, FlockTab checks it: known irreversible commands by rule, the rest by an AI judgment model. A stopped call waits for you: allow it once, always allow it, or keep it blocked, in the console or with tab guard. An agent stuck repeating itself can be paused.

Agent Guard in the agent's own session: ordinary work runs; a force-push, an unsandboxed ssh and an S3 delete are stopped.

Rules first, judgment second

A force-push, DROP TABLE, a production deploy, a payment or a recursive delete is stopped by rule. Everything else that can change something is judged by Agent Guard, and the reason is written in plain words.

As sensitive as you want

Pick how sensitive it is per agent: Relaxed, Balanced or Strict. Balanced is the default.

On Team and Fleet

Rules always run. Each Agent adds 5,000 Agent Guard checks a month, shared across your flock; past that, the rules keep working and the checks start again on the 1st.

It waits for you

The call is held while you decide. Allow it once and it runs; always allow it and that exact command is trusted on that Agent; keep it blocked and it stays refused. Every answer is recorded with who gave it.

Calls outside the project

ssh or scp to another machine, sudo, a database on another host, reading keys, or the harness turning its own sandbox off.

Your own patterns

Always flag "*prod*", always trust "ssh staging-bastion*": per Agent, in the console or with tab policy.

Secrets stay out

Keys, tokens and passwords are removed before a tool call is judged, and only a short summary is kept. Tool output never leaves the machine.

Also beyond the cap

Put a tab on every agent tonight.

Free for three agents. Install tab, set a cap, sleep.