60–90 second security demo

This reproducible terminal demo shows three outcomes: an allowed source read, a denied .env read, and a denied MCP write outside the workspace. It uses only temporary files and deletes them afterward.

Run it

git clone https://github.com/agent-util/agent-guard.git
cd agent-guard
npm ci
npm run demo

For a fast, unpaced verification:

npm run demo:quick

Suggested narration

Use the paced command while recording a terminal window.

1. 0–10 seconds: “Agent Guard is a policy and audit layer for supported AI-agent filesystem tools.”
2. 10–25 seconds: Show the agent reading src/app.ts. Explain that ordinary work inside approved paths continues normally.
3. 25–45 seconds: Show the .env attempt returning DENIED_PATH. Point out that the secret value never appears.
4. 45–65 seconds: Show the MCP write targeting ../outside-workspace.txt. Highlight Underlying MCP handler called: NO.
5. 65–80 seconds: Show the redacted audit entries and explain that decisions include agent, operation, reason, session, and correlation metadata without file contents.
6. 80–90 seconds: Close with: “Use Agent Guard with your agent’s native sandbox for defense in depth.”

Expected result

✅ ALLOW read src/app.ts
⛔ DENY read .env — reason: DENIED_PATH
⛔ DENY write ../outside-workspace.txt — reason: OUTSIDE_WORKSPACE
✅ Underlying MCP handler called: NO

The demo is an SDK/MCP enforcement demonstration. It does not claim to intercept arbitrary shell commands or unrelated processes.