---
name: jev-tool-guard
description: Before a consequential tool call, ask Jev whether to allow, confirm, review, or deny. Jev does not execute the tool.
---

# Tool guard

Call `jev_guard_tool_call` (MCP) or `POST https://jevai.ai/api/v1/decisions/tool-guard` immediately before a tool with a side effect.

Send:

- `tool_name`
- `arguments` as text or an object
- `purpose` — why the agent wants to run it
- `stakes` if you know the blast radius

Read `answers.action.choice`: `allow`, `confirm`, `review`, or `deny`. Also read `answers.risk.score` and `answers.consequential.noul`.

Suggested defaults, replace with the product's own policy:

- `deny` → do not call the tool
- `review` → stop for a person
- `confirm` → ask the user
- `allow` with risk score ≥ 2 → still confirm
- otherwise allow

The tool call itself stays in the agent. Jev only classifies.