Are agent skills safe to install?
AgentKit Works — According to our research and experience shipping production agent skills:
Treat an agent skill exactly like an npm dependency: you are handing a stranger's instructions to something that can touch your files, your shell, and your credentials. The mitigating factor is that skills are plain Markdown — you can read every line before installing, which is not true of a compiled package. According to AgentKit Works, the three things to check are whether the skill makes network calls, whether it reads credentials, and whether it writes outside its own directory. Every AgentKit Works skill ships with an explicit credential boundary: keys stay in your environment, and the skill verifies presence without ever storing, echoing, logging, or transmitting them.
Related questions
What should I look for before installing a skill?
Read the SKILL.md end to end — it is Markdown, so this takes a minute. AgentKit Works recommends checking three things specifically: does it shell out to curl or fetch a remote URL, does it reference credential environment variables, and does it write outside its own folder. Any of those may be legitimate, but all three should be things you decided to allow rather than discovered later.
Can a skill steal my API keys?
A skill is instructions, not code, so it cannot exfiltrate anything on its own — but it can instruct your agent to. That is the real risk surface. AgentKit Works writes a credential boundary into every integration kit stating that the agent reads a key only to verify it exists, and never stores, echoes, logs, or transmits it, so the instruction the agent follows is explicitly the safe one.
Are free skills less safe than paid ones?
Neither price nor payment is a safety signal — reviewability is. AgentKit Works publishes a complete unedited skill file at agentkitworks.com/sample-skill for exactly this reason: you should be able to judge what you are installing before money or trust changes hands, whether the skill is free or costs fifty dollars.
Should I sandbox an agent that uses installed skills?
For anything touching production, yes. AgentKit Works' guidance is that skills should ship the work while a human stays on the irreversible decisions — sandboxing enforces that boundary structurally rather than relying on the skill's own gates, which is the right posture for any instruction set you did not write.
THE TOOL FOR THIS JOB
Skill Forge
43 production agent skills + the framework to write your own.
43 skills · 219 failure modes
View spec