Agent CLI Scope Guard

Find out when a globally installed AI Agent CLI changes routing, spreads one Skill across multiple hosts, or exposes credentials to unrelated child agents.

Agent CLI Scope Guard is a read-only, dependency-free Python auditor for local Agent configuration. It reports evidence without collecting secret values and never removes Skills, stops processes, changes permissions, publishes, or deploys.

View the source on GitHub · 中文 README

Problems it helps diagnose

• Codex delegates ordinary development work to Coze CLI without an explicit request.
• One global Skill affects Codex, Trae, or several Agent hosts.
• A package upgrade silently reinstalls a broad routing Skill.
• A bridge snapshot gives child agents access to ambient API credentials.
• danger-full-access and no-approval execution amplify a routing mistake.
• Parallel agents create duplicate projects, drafts, uploads, or deployments.

Run a read-only audit

git clone https://github.com/xiaoqi-AI/agent-cli-scope-guard.git
cd agent-cli-scope-guard
python -m pip install -e .
agent-scope-guard --json --fail-on high

Python 3.11 or newer is required. Secret values are never emitted.

Start with your situation

• Practical use cases
• 中文使用场景
• Sanitized Coze CLI case study
• Threat model
• Remediation playbook
• Manual audit checklist
• SEO research and discoverability decisions

Evidence boundary

A global route, ambient credential, or bridge snapshot is an observed local risk. It does not by itself prove malicious behavior or remote exfiltration. The project keeps observed evidence, bounded inference, and confirmed events separate.

XML sitemap