# Semia

Semia is a security auditor for AI agent skills that reads a skill's Markdown and embedded commands as data, without ever executing it, and produces an evidence-backed report of the actions, effects and secret reads it may perform.

Agent skills are Markdown files that can embed shell commands, network calls and tool invocations, and they run with the user's own credentials, on the user's own machine, with the user's own data. Semia reads a skill's contents rather than running it and produces a report naming every action, effect and secret read it may exercise, each tied to the specific line of the skill that would cause it.

As a command-line tool it prepares a skill, synthesizes an analysis through a configured LLM provider, detects capabilities and writes a report, all in one command, with output saved under a per-run directory. It can also run as a plugin inside Codex, Claude Code or OpenClaw, where the host agent itself performs the synthesis step, so no separate API key is needed when used that way.

It is written in Python and released under the Apache license.

Website: https://github.com/berabuddies/Semia