| name | agents-security |
| description | (beta) Audit the user's local AI-agent setup with yeaboi: permission-bypass settings, wildcard allow rules, risky hooks, MCP server inventory, secret-shaped text and risky shell commands in session transcripts. Use when the user asks how safe their agent setup is, wants an agent security audit/scan, or asks about agent permissions or MCP server risks. |
Agent security workflows with yeaboi
Beta. These checks are deterministic pattern scans — an indicator, not
a security audit. A clean report means no known pattern matched, not that the
setup is safe; say so when you present it.
-
Run the scan with agents_security_scan. The default pass audits the
settings/MCP configs and any new or changed session transcripts; set
deep: true to re-scan every transcript (slower, thorough).
-
Present it worst-first: lead with the posture
(good / needs-attention / at-risk) and the critical/high findings, each with
its remediation. The mcp_servers table shows what is configured and its
risk flags (plain-http, unpinned-package, inline-credential).
-
Privacy is structural — findings carry pattern + file + line only.
Never ask the user to paste the matched secret back; point them at the
location and recommend rotation.
-
Compare over time with agents_security_history (newest first) — a
posture that regressed since the last scan is the headline.
-
Exports auto-save under ~/.yeaboi/exports/agentwatch/security/.
Error handling
Every tool returns {ok, llm_mode, warnings, data}. If ok is false, relay
error.message and its hint; don't retry blindly. llm_mode: "fallback"
means no LLM was reachable — the findings are still real, only the
summary/recommendations prose fell back to deterministic lines.
