Question 1

If Jev is down, does my session break or does the gate just stop gating?

Accepted Answer

The gate opens fully. Fail open means nothing is denied, forbid rules included, because a hook should never break a session. You outrank it too: a handler you name in the prompt is never denied, and only your own words are judged, since harness messages match passthrough and never replace the turn's verdict.

Question 2

Are the routing numbers from a real bench or just the examples?

Accepted Answer

Mostly the examples. The 24-skill marketing-team run listed at $0.000056 and 311 to 351 ms is illustrative, and the starter thresholds are docs illustrations rather than a measured bench. Build your own from a bench file of prompt and expect pairs, then run bin/gatekeeper bench to see which act threshold holds precision where you want it.

Question 3

Is this a skill router or a safety gate, which one am I installing?

Accepted Answer

Routing. It decides which agent or skill handles a prompt and never sends, publishes, deploys or spends. Six verdicts exist: allow, route, suggest, confirm, escalate and block. A fresh install is shadow mode and logs would_inject, advise adds context, and only enforce can block or deny a call.

Question 4

What leaves the machine on every prompt I type?

Accepted Answer

Up to 6,000 characters of each judged prompt go to the TypeSafe API under your key, plus the fields the rulebook lists in state and the roster's descriptions, in every mode except off including shadow. Your key and common secret shapes are masked first, but the README calls that best effort rather than a guarantee. Slash commands and plain acknowledgements are decided by code and never sent.

Question 5

How do I install and turn on Gatekeeper?

Accepted Answer

Clone the repo, pip install -e ".[test]", then run scripts/install_hooks.py --global --skill for a dry run and again with --apply to write hooks into ~/.claude/settings.json. A fresh install starts in shadow mode, which logs what it would do and changes nothing.

Question 6

What are Gatekeeper's four modes?

Accepted Answer

shadow judges prompts and logs would_inject without adding context or blocking; advise adds the route as context and never blocks; enforce may block or deny a call; off stops judging. Change mode with bin/gatekeeper mode agent-selection off, or set GATEKEEPER=off as a kill switch.

Question 7

What does a judged prompt cost?

Accepted Answer

Jev bills input tokens only. A judged prompt over a 24-skill roster costs about $0.00006, and over 143 skills and workflows about $0.0004. Gatekeeper sends the judged prompt, the rulebook's listed state fields and the roster descriptions to the TypeSafe API.

Gatekeeper

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community