Question 1

Is this just Claude Code's own auto-approval with extra steps?

Accepted Answer

It is explicitly not a replacement; the README tells you to use Claude Code's auto mode and Codex's Guardian. The difference is scope: jev-gate judges only the command and never the conversation, so nothing said in the session can argue it into allowing one. On 30 commands, auto mode let 8 of 12 dangerous ones run after the session claimed authorization; jev-gate let 0 run. It also catches mistakes, not a determined attacker.

Question 2

Will it start prompting me for commands I thought were safe?

Accepted Answer

Some. Safe commands ran with no prompt 9 of 12 times for jev-gate against 12 of 12 for Claude Code's auto mode, and it asks more often by design, at a median decision time of 0.28 s. Across 3,622 commands from the author's own history it approved 66% unaided and never approved rm, git push, git reset, or sudo.

Question 3

If the API is down or errors, does my whole session break?

Accepted Answer

No, it fails open. Any error, timeout, or missing key falls back to your harness's normal permission prompt, so a network blip never bricks a session. It also installs in observe mode by default and blocks nothing until you switch modes, and the mode persists in a file under your home directory so no restart is needed.

Question 4

Can I trust a fixed threshold, or will it drift when the model updates?

Accepted Answer

Expect drift. The README shows the same command answered ask at 0.78 on one run and deny at 0.94 on an earlier run, and says scores move when the model is retrained. That is exactly why thresholds live in your own config, why it ships a calibrate command, and why it tells you to read a week of logged decisions before turning anything on.

Question 5

What modes does the gate have, and how do I set them?

Accepted Answer

observe is the default and only logs; auto approves clear allows, blocks denies and asks about the rest, which is the one to run day to day; guard blocks denies and leaves every other prompt as it was; enforce blocks both denies and asks, for unattended runs. Set it with /jev-on, /jev-off or ./jev_gate.py --mode auto. The mode is saved in ~/.jev-gate/mode so no restart is needed, and JEV_GATE_MODE in the environment overrides it.

Question 6

What are the other two tools, and what do they cost?

Accepted Answer

route picks the model tier before you spend on the turn, stepping up one tier rather than to the top when confidence falls below 0.5. rank orders up to 255 options against one criterion in a single call. keep decides which pieces of context still earn their place, keeping each item verbatim or dropping it, because a summary silently loses the exact path or error. All three are CLI commands and MCP tools, so an agent can call route_turn, rank_options and keep_context itself.

Question 7

What does it cost, and how do I verify the setup?

Accepted Answer

A dangerous command is denied in 371 ms for $0.0000189, roughly a cent a day at 500 checks. After installing the plugin and adding OPENROUTER_API_KEY to ~/.claude/settings.json, work for ten minutes while it only watches, run /jev-status to see what it would have approved and blocked, then /jev-on. Read ~/.jev-gate/decisions.jsonl for a week first: that file is the only source of thresholds that will fit your work, and /jev-calibrate derives them for you.

Jev Gate — eugeniughelbur

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community