Agent Guardrail

It is an ultra-low-latency safety guardrail for autonomous coding agents such as Claude Code, OpenAI Codex, Google Antigravity, Cursor and Grok Build. It intercepts shell commands, script executions, file deletions, git operations and sensitive directory changes before they run.

Jev is one of the selectable Tier 2 neural decision engines, alongside Liquid AI d1-3B, CLM and Pure Rules. The README's engine comparison measures TypeSafe AI Jev at about 60.0 ms average latency, evaluating contextual risk and ambiguous commands.

Tier 1 is a deterministic shield of compiled regex and path policy that blocks destructive patterns in under 0.1 ms. Tier 2 is a neural decision engine for ambiguous commands. Tier 3 routes rare or esoteric commands to a local Qwen 3.8B that explains side effects and feeds the analysis back, lifting confidence from 34% to over 95%.

The README lists Claude Code, OpenAI Codex / ChatGPT, Google Antigravity, Cursor and Grok Build, with a lifecycle hook bridge, an MCP server integration, a one-command installer and per-platform setup sections; sessions activate through a /guardrail command.

The project states token cost of $0.00 with local engines and advertises 100% offline, local and air-gapped operation when using its local models. The cloud LLM guardrails it compares against are reported at $2.40 to $7.20 per 1,000 actions.

Ten-plus defenses: recursive wrapper unwrapping to five levels, GTFOBins and CLI argument-injection blocks, env preload detection, staged payload inspection, local script peeking, anti-self-tampering, reverse-shell and exfiltration blocks, shell-profile poisoning prevention, payload decoding and homoglyph normalization.

The GitHub user Bl4ckk01 publishes it under the MIT license; the repository was created 2026-10-08 and pushed 2026-10-17, and its badge set covers Python 3.9-3.12 with 135 passing tests.