User repository on GitHub
A Classifier for AI Agent Shell Commands, Powered by Decision Models or LLMs
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Integrations & Adapters
Tiny Bouncer
A classifier that screens shell commands requested by LLM agents before they run: each command goes to an external judgment backend — TypeSafe's Jev is the production default — and the verdict becomes allow, block, or the normal interactive prompt in the OpenCode V2 harness.
- Stars
- 0
- Language
- Go
- Reading time
- 2 min
- Published by
- Community
- Author
- radupopescu
- Last verified
- Oct 2026
Highlights
- Screens LLM agent shell commands before execution and turns the verdict into allow, block or the interactive prompt in OpenCode V2.
- TypeSafe's Jev is the production default backend and the only one that meets the project's hard FNR = 0 safety gate.
- Fail-safe design: if the classifier is unreachable, times out or misbehaves, the agent is asked interactively, never silently auto-run.
- The classifier only raises strictness; commands explicitly denied in opencode.json(c) stay denied and it never grants permission on its own.
- v0.1.0 status: all planned work complete, with decider, api and afm backends documented as comparison-only.
Watch out
README and GitHub metadata reviewed 2026-10-17; the captured text cuts off at 'How it', so setup instructions, the measured findings in doc/findings.md and the exact FNR figures are not visible here. The README also notes the manual OpenCode smoke test had not been run.
Questions about this entry
Short answers drawn from the linked project or page.
A classifier that screens shell commands requested by LLM agents before they run, sending each command to a judgment backend and turning the verdict into a permission decision in the OpenCode V2 harness.
TypeSafe's Jev by default; the README says Jev is the production default and the only backend that meets the hard FNR = 0 safety gate, while the optional decider, api and afm backends are comparison-only.
It fails safe to ask: if the classifier is unreachable, times out or misbehaves, the agent is asked interactively rather than silently auto-running the command.
No. The classifier only raises strictness: commands explicitly denied in your opencode.json(c) stay denied, and it never grants permission on its own.
The backend-agnostic contract returns allow, deny or ask, so other judgment backends can be added without touching the CLI contract or the OpenCode plugin.
v0.1.0 with all planned work complete per the README; the GitHub metadata lists no licence, and the manual OpenCode smoke test had not yet been run.
The README points to doc/findings.md for measured results and doc/plan.md for the plan; the digest truncates the README at 'How it', so the numbers themselves are not captured here.
More Jev resources like this
An opencode plugin (also for pi and Claude Code) that puts every bash command in front of a System One decision model before it runs: a small model scores destructiveness, secret leakage, and your repo's GUARDRAILS.md policy and blocks bad calls.
Jev Integrations & Adapters#community#github#opencode
Community
berget-ai/guardrails-mdTypeScript
berget-ai/guardrails-mdTypeScriptAn auto-mode plugin for the Pi coding agent that classifies shell commands before they run: built-in rules decide most calls, and Jev (hosted) or Kev and Laya (local, CPU, offline) route the rest to run, confirm, or block.
Jev Integrations & Adapters#community#github#pi
A Kubernetes admission webhook that asks Jev to allow, flag, or deny a manifest change for intent-shaped policies a Rego rule cannot express, keeping hard rules in Rego and failing closed by default.
Jev Integrations & Adapters#community#go#kubernetes
Community
An LLM-agnostic guardrail API built on Jev: it checks prompts and responses against YAML policies and returns PASSED, BLOCKED or FAILED. It sits between your app and any LLM provider, so guardrail decisions do not depend on the provider you call.
Jev Integrations & Adapters#community#github#guardrails
Community
alpkeskin/jev-guardrailGo
alpkeskin/jev-guardrailGoA fork of Ollama that adds POST /v1/systemone: the endpoint is forwarded to a custom llama-server from nace-ai/llama.cpp on the edlm branch, which serves Drex DLM. The fork is MIT licensed and Go-based like upstream Ollama.
Jev Integrations & Adapters#community#github#ollama
Community
nace-ai/ollamaGo
nace-ai/ollamaGoA Unix-pipeable CLI that asks Jev (or Berget System One models) one question over stdin, with shell-safety exit codes and a calibrate subcommand.
Jev Integrations & Adapters#community#cli#shell
Community
frodi-karlsson/onesieGo
frodi-karlsson/onesieGoFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.