Skip to content
SaferSkills Docs

Managing Your Agents

Most developers run more than one agent — Claude Code here, Cursor there, a CLI agent for scripts — and each keeps its own skills, MCP servers, and hooks in its own directory. The saferskills CLI consolidates all of them under one view: saferskills list enumerates what is installed across every agent it detects, and saferskills capability (with no target) audits the whole installed inventory in one pass. Nothing is verified by trust; everything is scannable.

Why do my agents need a single view?

Section titled “Why do my agents need a single view?”

Capabilities scatter by design. Claude Code loads skills from ~/.claude/skills/, Cursor reads MCP servers from ~/.cursor/mcp.json, Codex uses ~/.codex/skills/, and five more agents each have their own paths. Without a consolidating tool you have no single place to answer the basic question: what is installed, where, and is any of it risky? A poisoned MCP server or an auto-executing hook does not announce itself — and the Sonatype 2024 supply-chain report found a 156% year-over-year rise in malicious open-source packages, so the inventory is worth knowing.

How do I list what is installed?

Section titled “How do I list what is installed?”

saferskills list enumerates the capabilities installed across the agents the CLI detects, in one table. It reports each item’s kind (skill, MCP server, hook, plugin) and the agent it belongs to, so you see your whole footprint without opening eight config files.

Terminal window
saferskills list

Add --json for machine-readable output on stdout (human text always goes to stderr). The full reference is on the list command page.

How do I audit everything at once?

Section titled “How do I audit everything at once?”

saferskills capability with no target runs a static scan of every installed capability the CLI can find — the same deterministic engine that scores a repo, pointed at your local installs. Each capability gets a 0–100 aggregate score and a finding list, so a risky item surfaces with its rule_id and evidence rather than a vague warning.

Terminal window
saferskills capability

Point capability at a specific path or URL to scan one artifact instead. See the scan command page (which documents the capability command) for targets and flags.

How does this fit with installing and scanning?

Section titled “How does this fit with installing and scanning?”

The audit is the read side of the same loop that governs installs. When you install with saferskills install <name>, the CLI gates on the aggregate score — the default minimum is 90 (SAFERSKILLS_MIN_SCORE) — so a low-scoring capability is held back unless you choose otherwise. Auditing with saferskills capability then lets you re-check what is already installed against the current rubric.

Where do I go next?

Section titled “Where do I go next?”