jev-gate approval firewall

Ask a hosted judge before risky agent actions. Returns approve, hold, or escalate with a confidence number and the rule that fired, and keeps a reviewable decision log. Use before deleting data, spending money, touching production, or messaging customers.

jev-gate — ask before you act

Run this before an action you would not want to take unsupervised: deleting or overwriting data, moving money, touching production, messaging a customer, changing credentials. The gate answers with a verdict, a confidence number, and the rule that fired — and keeps the decision in a log a human can review later.

Setup (once)

1. Get a key at https://jev-gate.fly.dev — 14-day free trial, card required and not charged during the trial; cancel any time in the billing portal.
2. Export it:

bash

export JEV_GATE_KEY=jgate_...

That is the whole install: one key, one HTTPS call. No package, no dependencies beyond curl.

Check an action

bash

scripts/jev-gate-check "about to delete 1,204 rows from the staging events table"

Exit codes, which is all an agent needs:

code	verdict	what to do
0	approve	continue with the action
2	hold	pause this action and queue it for routine human review; do not retry
3	escalate	stop, tell the user a human decision is required
1	error	the gate could not answer — treat as blocked, never guess

The raw call, if you would rather do it yourself:

bash

curl -s https://jev-gate.fly.dev/v1/gate \
  -H "Authorization: Bearer $JEV_GATE_KEY" \
  -H 'content-type: application/json' \
  -d '{"transcript":"agent is about to drop the orders table in production","tool_calls":["psql drop"]}'

json

{
  "action": "escalate",
  "confidence": 0.38,
  "violation": 0.71, "needs_human": 0.72, "risk": 2.4,
  "reasons": ["risk 2.40 ≥ 2.00", "needs_human 0.72 > 0.60"],
  "latency_ms": 474
}

Branch on action — never on the prose. reasons is the rule that fired, and it is what makes the decision explainable to the human who reviews it.

Reading the log

Every call lands in the account's dashboard at https://jev-gate.fly.dev/dashboard with the transcript, the verdict, the reasons, and the latency. There you can tune the four thresholds it branches on (approve if confidence ≥, hold if violation >, escalate if needs_human >, escalate if risk ≥), work the review queue for the muddy middle, and export CSV/JSON. Reviews are closeable: mark reviewed, annotate, reopen.

Going further

• MCP: curl -sO https://jev-gate.fly.dev/client/jev_gate_mcp.py, then point your MCP config at that file — the gate becomes a check_action tool.
• Python, any framework: from jev_gate_mcp import guarded wraps your tool dispatcher so held and escalated calls raise before the side effect.
• Claude Code: curl -sO https://jev-gate.fly.dev/client/claude_code_hook.py and wire it as a PreToolUse hook; exit 2 blocks the call.
• All client files: https://jev-gate.fly.dev/client

Rules of thumb

• Gate the action, not the thought: pass what is about to happen and the context behind it.
• A hold is not a failure — it is the product working. Queue it and keep moving.
• If the gate is unreachable, block. A firewall that guesses is not a firewall.