# AI use cases for Jev Guard (leepokai/jev-guard)

Repo: [leepokai/jev-guard](https://github.com/leepokai/jev-guard) | Use cases: 5 | Page: https://dedupedai.com/use-cases/leepokai-jev-guard/

jev-guard is a security hook for coding agents that scores each tool call with session context and returns deny, ask, or allow, flags prompt injection in tool results, and checks skills and plugins, powered by Jev.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Guard Claude Code tool calls with deny and ask decisions: https://dedupedai.com/use-cases/leepokai-jev-guard/guard-claude-code-tool-calls/
- Scan agent skills and plugins for unexpected behavior: https://dedupedai.com/use-cases/leepokai-jev-guard/scan-skills-and-plugins-before-install/
- Add a guard layer to Codex through its plugin browser: https://dedupedai.com/use-cases/leepokai-jev-guard/guard-codex-with-hook-trust/
- Require approval for risky Cursor shell and MCP calls: https://dedupedai.com/use-cases/leepokai-jev-guard/guard-cursor-shell-and-mcp-calls/
- Put a guard proxy between Zed and an ACP agent: https://dedupedai.com/use-cases/leepokai-jev-guard/guard-acp-editor-agent-in-zed/

## How to guard Claude Code tool calls with deny and ask decisions with Jev Guard

Repo: [leepokai/jev-guard](https://github.com/leepokai/jev-guard) | Difficulty: Intermediate | Time: 15 minutes | Basis: From the README

Add a pre-tool check to Claude Code that blocks destructive commands and asks before risky ones.

**Who it is for:** A developer using Claude Code on a real project who wants auto-style approvals without trusting every command the agent proposes.

**The problem:** Claude Code can run shell commands and edits on its own, and a single destructive or off-task command can damage a workspace before anyone reviews it.

**The outcome:** Each Claude Code tool call is scored with the session context, destructive calls are denied, risky calls show an approval prompt, and results are scanned for injected instructions.

### Steps

1. Run `npm i -g jev-guard`
2. Run `jev-guard key "…"` with your TypeSafe, Vercel AI Gateway or OpenRouter key
3. Run `jev-guard install claude`
4. Start Claude Code and run a harmless command, then run `jev-guard check Bash '{"command":"rm -rf ~/"}'` to confirm the deny output

### Prompt to paste into your agent

```text
Install jev-guard from the leepokai/jev-guard repo for Claude Code on this machine, then run `jev-guard check Bash '{"command":"rm -rf ~/"}'` and report whether the output starts with DENY. Do not run the destructive command itself.
```

### Caveat

The README says the guard calls a Jev model for each tool call, so you need a provider key and network access before any decision is made.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/leepokai-jev-guard/guard-claude-code-tool-calls/

## How to scan agent skills and plugins for unexpected behavior with Jev Guard

Repo: [leepokai/jev-guard](https://github.com/leepokai/jev-guard) | Difficulty: Intermediate | Time: 15 minutes | Basis: From the README

Check installed skills, plugins and instruction files for exfiltration, covert execution or overrides.

**Who it is for:** A developer who installs community skills or plugins for a coding agent and wants to check them before the agent obeys them.

**The problem:** Skills, plugins and CLAUDE.md or AGENTS.md files tell the agent what to do, so a malicious one can quietly exfiltrate data or override other rules.

**The outcome:** Every installed skill or plugin is checked for behavior its installer would not expect, and a verdict is produced on demand.

### Steps

1. Run `npm i -g jev-guard`
2. Run `jev-guard key "…"` with your provider key
3. Run `jev-guard scan-skills` to check installed skills and plugins
4. Review any flagged file by hand before the agent loads it again

### Prompt to paste into your agent

```text
Use the leepokai/jev-guard repo to scan the skills and plugins on this machine by running `jev-guard scan-skills`. Report every flagged file with its kind, and confirm that no file was modified by the scan.
```

### Caveat

The README describes the scan categories but not the exact output format, so check the scan report yourself rather than assuming a flag means a confirmed attack.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/leepokai-jev-guard/scan-skills-and-plugins-before-install/

## How to add a guard layer to Codex through its plugin browser with Jev Guard

Repo: [leepokai/jev-guard](https://github.com/leepokai/jev-guard) | Difficulty: Intermediate | Time: 15 minutes | Basis: From the README

Install the Jev Guard plugin in Codex and trust its hooks so tool calls are checked.

**Who it is for:** A Codex user who wants the same deny and warning checks as other agents, installed from inside Codex.

**The problem:** Codex runs tool calls without a second review, and the user needs a guard that installs through the plugin system rather than manual config editing.

**The outcome:** Codex tool calls are checked before they run, with deny decisions and warnings for ask-level risks, and results are flagged when they contain injected text.

### Steps

1. Run `codex plugin marketplace add leepokai/jev-guard`
2. Open the plugin browser in Codex and install Jev Guard
3. Run `/hooks` in Codex and trust the Jev Guard hooks
4. Trigger a harmless tool call and confirm the guard runs

### Prompt to paste into your agent

```text
Install the jev-guard plugin in Codex from the leepokai/jev-guard marketplace, trust its hooks with /hooks, and report whether the hooks list shows jev-guard as trusted. Do not run any destructive command to test it.
```

### Caveat

The README says Codex has no ask decision yet, so ask-level risks appear as a warning rather than a prompt.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/leepokai-jev-guard/guard-codex-with-hook-trust/

## How to require approval for risky Cursor shell and MCP calls with Jev Guard

Repo: [leepokai/jev-guard](https://github.com/leepokai/jev-guard) | Difficulty: Intermediate | Time: 15 minutes | Basis: From the README

Install the Cursor guard so shell and MCP tool calls are denied or sent for approval.

**Who it is for:** A Cursor user who lets the agent run shell commands and MCP tools and wants a check before risky ones execute.

**The problem:** Cursor agents can run shell and MCP calls that touch files or external services, and a single bad call can cause real damage.

**The outcome:** Destructive shell and MCP calls are denied, risky ones ask for approval, and tool results are scanned for injected instructions.

### Steps

1. Run `npm i -g jev-guard`
2. Run `jev-guard key "…"` with your provider key
3. Run `jev-guard install cursor`
4. Restart Cursor and trigger a shell call to confirm the approval prompt appears

### Prompt to paste into your agent

```text
Install jev-guard for Cursor on this machine with `jev-guard install cursor` and report the path of the hooks file it wrote. Do not run any shell command that changes files.
```

### Caveat

The README says Cursor's preToolUse cannot ask, so the ask decision applies only to shell and MCP calls, not every tool.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/leepokai-jev-guard/guard-cursor-shell-and-mcp-calls/

## How to put a guard proxy between Zed and an ACP agent with Jev Guard

Repo: [leepokai/jev-guard](https://github.com/leepokai/jev-guard) | Difficulty: Advanced | Time: 30 minutes | Basis: From the README

Wrap an ACP agent in Jev Guard so permission requests cover terminal and file writes.

**Who it is for:** A developer who runs an ACP-compatible coding agent inside Zed and wants approval requests for terminal and file writes.

**The problem:** An ACP agent can ask to create terminals or write files through the editor, and the editor alone does not score each call for risk or injected content.

**The outcome:** Terminal creation and file writes raise a permission request, and reads and terminal output are flagged when they contain injected text.

### Steps

1. Run `npm i -g jev-guard`
2. Run `jev-guard key "…"` with your provider key
3. In Zed settings, add an agent_servers entry whose command is node and whose args include the path to src/cli.js, then `acp`, `--`, and your agent command
4. Start a session in Zed and confirm a permission request appears for a terminal or file write

### Prompt to paste into your agent

```text
Configure a Zed agent_servers entry in <settings-file> that runs the jev-guard acp proxy around <agent-command>, following the leepokai/jev-guard ACP example. Report the final JSON and confirm the proxy is the command that starts the agent.
```

### Caveat

The README says the proxy only sees traffic that passes through the client, so tools the agent runs on its own are not covered.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/leepokai-jev-guard/guard-acp-editor-agent-in-zed/