What it stops
Known irreversible commands are stopped by rule: a force-push, DROP TABLE, deleting outside build folders, a production deploy, a payment, sending an email. Everything else that can change something goes to TypeSafe's judgment model, which says how sure it is that the call cannot be undone. Read-only calls are not checked at all, and a tool FlockTab does not know counts as a write.
Allowing a call means saying nothing: the harness's own permission prompts still apply as they always do. Agent Guard only adds a stop on top.
Turn it on
- Open Control, Agent Guard and pick a mode for each Agent, or one for all of them. Each change saves at once and applies from the Agent's next launch.
- To start every new Agent the same way, set Agent Guard for new Agents in Settings.
- Start in Watch for a day, look at what it would have stopped, then move to Ask first or Deny.
tab policy my-agent --guard ask # one Agent
tab policy --all --guard watch # every Agent at once
tab policy my-agent --guard-level strictThe four modes
| Mode | What happens to a call it flags |
|---|---|
| Off | Nothing is checked. The default. |
| Watch | Judged and recorded, never stopped. Use it to see what the guard would catch. |
| Ask first | You confirm it in the terminal before it runs. |
| Deny | Refused, and the agent is told why. For Agents nobody is watching. |
How sensitive
Sensitivity decides what counts as a call to stop. Balanced is the default.
| Level | What it stops |
|---|---|
| Relaxed | Stops only what surely cannot be undone, by rule: a force-push, deleting outside build folders, dropping a database, a production deploy, a payment, an email. Agent Guard flags a call only when it is 80% sure it cannot be undone. For trusted agents you watch anyway. |
| Balanced | Everything Relaxed stops, calls Agent Guard is 50% sure cannot be undone, and calls outside the project: ssh, sudo, credentials, a database on another host, the sandbox off. Everyday project work rarely asks. |
| Strict | Also any git push, any delete, a web request that sends data, a database or cloud write, and an MCP tool that creates, changes or sends. Agent Guard flags from 30%. For new or unattended agents, or anything near production. |
Calls outside the project
With the guard on, it also stops calls outside the project (ssh, sudo, keys): ssh, scp or rsync to another machine, sudo, kubectl exec, a database on another host, reading private keys or credentials files, or the harness turning its own sandbox off. Watch records these calls, Ask holds them until you confirm, Deny refuses them. It is on unless you turn it off (tab policy my-agent --reach off, or the checkbox beside the mode), and Relaxed never checks it. When only the model thinks a call leaves the folder (a grep or a cd elsewhere), it asks only if the call may also be hard to undo.
Messages to other agents
With the guard on, a message one agent sends to another agent or AI is checked on your machine before it leaves: a Claude Code SendMessage to another session, a Codex hand-off to another agent, a call to the Codex MCP server, and a shell command that calls an AI provider's API. This check is stricter than the one on other calls. It refuses test keys too, a value named as a credential however short (githubToken=ab12), a key split by spaces or new lines, and any long random-looking string. A sentence after a credential's name is not a value (SENTRY_AUTH_TOKEN: removed from the guard), unless a word in it has a digit or mixed case, or the words are in quotes. A full 40-character commit SHA passes when it is a commit in the session's repository; any other one is refused: send a short SHA (8 to 12 characters) instead. In every mode but off, a message that carries one is refused, never asked about, since a sent secret cannot be unsent. A message the guard cannot read, such as one over 1 MiB, is refused too.
The refusal names the kind of secret and who the message was for, never the value; a recipient name that looks like a secret itself shows as “another agent”. Only that, and the message's length, are kept; the message itself is not stored. A session's own subagents, started with the Agent tool, are not checked this way. A SendMessage is, even between a subagent and its lead, since its address cannot tell a teammate from another session.
In Claude Code, a reply from the Codex MCP server reaches your agent marked as data from another agent, never as instructions. A reply to a SendMessage arrives as Claude Code's own message, and Codex cannot have a reply replaced, so neither is marked.
Your own rules
Tune it per Agent with patterns. A flagged pattern is always stopped; a trusted one always goes through, before every rule. Case and repeated spaces do not matter, and * matches anything. A trusted pattern with * never matches text that chains or redirects (; && | > and the rest), so npm test* never trusts npm test; rm -rf ~. Up to 40 of each.
tab policy my-agent --flag "*prod*"
tab policy my-agent --trust "ssh staging-bastion*"
tab policy my-agent --forget "*prod*"When a call is held
In Deny (and in Ask first where the harness cannot ask), a flagged call waits 45 seconds for a person. Allow it in that time and it runs there and then; otherwise it is refused and stays on the list.
tab guard # what is waiting, and what was blocked
tab guard allow <id> # propose letting it run once: confirm in the console
tab guard allow <id> --always # also trust that exact command on this Agent
tab guard deny <id> # keep it blockedAn allow from the terminal is only a proposal: it prints a link and a code, and the call is allowed when you confirm it in the console, signed in. The agent can reach its own terminal, but not your browser session. The Review tab of Agent Guard lists every held call and what people decided.
When the guard itself fails
If Agent Guard's own check fails on a call (an error, or no answer in a few seconds), the call does not run unchecked. In Ask first it asks you, and in Deny it is refused, with the reason "Agent Guard hit an error checking this call". Watch and Off let it run, as always. Plain reads of this machine (reading, searching and listing files) still run. Web fetches, messages and every other tool get the mode's answer. To let an Agent work while a fix ships, switch it to Watch. Each failure is listed on the Agent with the error's name, and FlockTab is told about the error, never about the call.
The same goes for a check that takes too long. Your harness runs a call whose hook took longer than its time limit, so the check stops itself first, after a few seconds, and answers "Agent Guard ran out of time checking this call". A tool call larger than 1 MB is not checked at all and gets the same answer, with "Agent Guard can't check a tool call over 1 MB": split it into smaller calls.
On macOS and Linux a small watchdog backs this up: if the check is still stuck a few seconds later, the watchdog stops it and gives the same "ran out of time" answer itself, well before your harness's time limit, so a stuck check never lets the call run. On Windows there is no watchdog. There, the check's own time limit is what stops it.
Learned from your answers
Say no to Agent Guard's question in Claude Code and it flags that exact command on that Agent at once. Say yes and nothing changes until you trust it on the Learned tab of Agent Guard. Your own flagged and trusted patterns are never touched. After 40 learned rules an Agent stops learning until you undo some.
The loop watch
A separate switch on the same page. It looks at the last ten tool calls every three calls and spots an agent repeating itself without progress, or polling the same check on a timer. Watch records it; pause also closes the tab, the same as tab close, once it has been stuck two checks in a row (--pause-after 1-5 changes that).
tab policy my-agent --loops pause --pause-after 3What leaves your machine
- Each checked call (the command and its arguments, including the text of a file write) goes to TypeSafe's judgment model with keys, tokens and passwords removed. Only a 200-character summary is kept.
- Tool output never leaves the machine.
- If flocktab.com cannot be reached, the rules still run on your machine; only the judgment is skipped.
Which agents
tab claude, tab codex, tab gemini, tab agy, tab opencode, tab goose, tab cline, tab grok, tab kimi and tab pi. Codex, OpenCode, Goose, Cline and Kimi Code cannot ask from a hook, and nobody can answer a headless run such as -p: there the agent is told to ask you instead. Codex asks once to trust FlockTab's hooks (choose Trust all and continue); until then its calls run unchecked, and tab says so.
Plans and limits
| Plan | Agent Guard |
|---|---|
| Solo | Not included |
| Team | Rules always, and 5,000 checks a month for each Agent, pooled, up to 500,000 |
| Fleet | Rules always, and 5,000 checks a month for each Agent, pooled: at least 200,000, up to 500,000 |
Each Agent adds 5,000 Agent Guard checks a month, shared across your flock. An Agent counts while it holds a seat, the same as on your bill: an archived one until its seat frees a day later. The guard and the loop watch share the month's checks (calendar month, UTC). Past them the rules keep working and the checks resume on the 1st; every console page says so while it lasts.