MCP tool
Registers as an MCP server in Cursor, Claude Code, and Windsurf. Agents call check_package before installing and get a verdict plus alternatives.
Shell hook
A Cursor beforeShellExecution hook blocks raw npm/pip/cargo installs. The agent must retry through scanopy — it cannot bypass the gate.
Skills & agent context
Scans SKILL.md, CLAUDE.md, hooks, and MCP configs for hidden Unicode, exfil directives, shell execution, and API base-URL hijacks — with audit or block policy.
Set up in one command
# Detect installed AI tools and lock them to scanopy
scanopy ai-protect
# Scan skills and instruction files for malice
scanopy skills
# Configure every supported tool, even if not detected yet
scanopy ai-protect --allThe installer runs ai-protect for you automatically. Supported today: Cursor, Claude Code, OpenAI Codex, Gemini CLI, Windsurf and Aider. Turn on skills scanning in Policy → AI skills & agent context.
What the agent sees when something is blocked
✘ BLOCKED reqests@2.0.0
possible typosquat of "requests"
→ safe alternative: requests
scanopy pip install requests
$ scanopy skills
✘ SKILL-001 zero-width Unicode in CLAUDE.md
✘ SKILL-004 ANTHROPIC_BASE_URL hijackPackages and agent instruction files are both in scope. Malware doesn't get a free pass just because it arrived as a skill instead of an npm package.