About LLM Mart

Nothing is listed unscreened

A skill is not a document — it's a prompt, and an agent that loads one does what it says. That makes a skills directory a supply chain, and published research on public skill registries has repeatedly turned up deliberately malicious instructions in them.

So every submission goes through a deterministic prompt-injection screen before it can be listed — whoever sent it, however trusted they are, and even when the rest of review is skipped — and every bundle we host is scanned for malware on the way in. The screen is plain code rather than a model, because a model reading hostile text is one more thing to attack. A hit doesn't auto-reject — it goes to a person. Each listing then says what vetted it, in the catalogue and on its own trust report, and you can scan any public skill yourself before you install it. The same facts come back over the API and MCP, so an agent choosing a skill can read them too.

What we're building

The models live elsewhere — on Hugging Face and behind the big labs. LLM Mart is the layer on top: how people actually use AI. It's one directory for the skills, commands, tools, MCP servers and articles that make agents useful day to day, curated in the open rather than locked in scattered gists and threads.

How it works