agentguard

Checks whether what you already have installed in Claude is safe. Say: 'Scan my Claude setup', 'What does EXFIL-001 mean?', 'Clean up my duplicate skills', 'Set up AgentGuard'. Audits an agent environment with AgentGuard (`aguard scan`): ~/.claude or a project's .claude, for prompt injection, credential exfiltration, arbitrary-execution grants, silent hooks and malicious skills/MCP servers/subagents/CLAUDE.md, then triages findings into a fix plan; also junk cleanup (`aguard clean`). Trigger when the user asks to scan, audit or check their agent setup or whether installed skills are safe, asks what a rule ID means (INJ-001, EXEC-001, EXFIL-002, PERM-006, HOOK-001, SUP-004, COV-000, GATE-001, ...) or why their score is what it is, wants config cleaned up, asks how to use AgentGuard (references/usage.md, no scan), wants the deeper AI check or to set up/test the LLM judge (references/llm.md), or asks in any wording to set up, install or get started with AgentGuard (references/setup.md).

Turns on automatic protection: every skill is checked before it loads, and anything with a finding asks you first. Say: 'Turn on the gate', 'I trust this skill, stop asking'. Install, inspect, repair or remove AgentGuard's load-time gate (`aguard hook`) — a Claude Code hook that audits a skill before an agent loads it and asks the user about anything carrying a finding, and manage the hash-keyed approvals it remembers (`aguard approve` / `approvals` / `forget`). Trigger when the user wants automatic protection rather than remembering to scan, asks to install/uninstall the aguard hook or set up hook-based scanning of skills, asks why the gate is prompting (or has stopped prompting), asks to trust or untrust a skill so it stops asking, or hits a GATE-000 / GATE-001 finding.

Checks a new skill before you install it. Say: 'I downloaded this skill, is it safe to install?' and hand over the folder or the link. Vet a single AI agent artifact with AgentGuard (`aguard check`) before it is installed, committed or trusted — one skill directory, plugin, MCP config, subagent, slash command, hook script, CLAUDE.md or loose file. Exits non-zero at or above a severity threshold, so it also drops into a git pre-commit hook or CI. Trigger when the user asks whether a specific skill/plugin/repo is safe to install, pastes or links one and asks you to review it, asks to check a downloaded or third-party skill before adding it to ~/.claude, wants a security gate in CI or pre-commit for the skills they publish, or wants an artifact's canonical hash.