How does this compare to asking an LLM whether a call looks safe?

An LLM judge returns free-form text that can be misread. Jev returns structured Choice, Score, and Noul with calibrated probabilities, so routing code can branch deterministically.

What happens on a write to a sensitive path?

The documented flow sends state plus request to Jev, receives a rejected choice with high score, and the guardrail blocks the call, logs the attempt, and notifies the agent.

How do I try it without wiring my own agent first?

Run demo.py for the quick path, streamlit_app.py for the dashboard on localhost, or the simple_agent and file_operations_agent examples for shaped tool flows.

How do I tune it for different tool types?

Adjust confidence thresholds per tool type, add domain state to Jev inputs, add custom fallback for noul responses, and add audit logging and alerting.

How do I install and configure it?

Run pip install -r requirements.txt, then set JEV_API_KEY in config.py or as an environment variable with export JEV_API_KEY=... . The core ships as guardrail.py, jev_client.py and config.py, with demo.py, a Streamlit dashboard and two example agents alongside them.

What exactly does the guardrail decide on?

Jev returns a Choice of Approved, Rejected or Suggested Alternative, a 0 to 1 Score for confidence and a Noul for uncertainty. The guardrail allows execution on an approved high-confidence result, blocks rejected or low-confidence ones, surfaces alternatives when offered, and requests clarification on a Noul.

How do I wire it into my own agent?

Build a ToolCallRequest with tool_name, parameters, agent_state and context, then call guardrail.execute_tool_call(request, your_executor). A blocked call raises PermissionError, so your except branch can catch it and hand the agent a safer option.

CommunityJev Agents & MCP

Jev Agent Guardrail

Python guardrail that intercepts agent tool calls, sends agent state plus tool details to Jev, and allows, blocks, or reroutes execution based on Choice, Score, and Noul answers.

See highlights

Open github.com

Stars: 1
Language: Python
Published by: Community
Author: bhanuprakashk90-tech
Last verified: Oct 2026

bhanuprakashk90-tech/jev-agent-guardrail

User repository on GitHub

View on GitHub

A real-time safety guardrail and interactive monitoring dashboard for AI agent tool calls powered by TypeSafe AI's Jev model."

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

This project implements a safety guardrail for AI agents using Jev (TypeSafe AI's System One Model) to validate tool calls before execution. Jev is a specialized AI model that provides typed, structured decisions (Choice, Score, or Noul) from application state, enabling safe routing, risk checks, and tool-call guardrails without replacing LLMs. Intercept Tool Calls: When an AI agent wants to execute a tool, the request is intercepted. State Analysis: Current agent state + proposed tool details are sent to Jev API.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Maps Choice to approved or rejected, Score to confidence, and Noul to uncertainty.
Ships guardrail.py, jev_client.py, config.py, demo.py, and a Streamlit dashboard.
Includes simple_agent and file_operations_agent examples plus ToolCallRequest usage.
Blocked calls raise PermissionError while uncertain calls request clarification.

Watch out

Needs a Jev API key via config.py or environment; thresholds and fallback logic need tuning per tool. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

agent-guard

Claude Code PreToolUse hook that screens Bash, file edits, and tool calls with Jev risk, policy, and allow-ask-deny judgments before execution.