Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/openagen/zeroclaw/llms.txt

Use this file to discover all available pages before exploring further.

Skills package agent behaviours as portable, auditable units: a TOML manifest that declares metadata and a SKILL.md file that contains the instructions injected into the agent’s system context. No Rust compilation is required — skills extend agent behaviour entirely through configuration and prompt engineering.

​
What a skill contains

A skill directory contains at minimum two files: 
my-skill/
├── skill.toml    # manifest: name, version, description, author, tags
└── SKILL.md      # instruction content injected into the agent's context
The manifest declares identity and metadata. The SKILL.md file is what the agent actually reads — write it as a set of instructions, references, or workflow steps the agent should follow when the skill is active.

​
Managing skills

zeroclaw skills list            # list installed skills
zeroclaw skills install <path>  # install a skill from a local directory or archive
zeroclaw skills remove <name>   # uninstall a skill by name
zeroclaw skills audit <path>    # run the security audit gate without installing
zeroclaw skills audit runs the same checks as install but exits without writing anything. Use it to inspect a skill before committing to installation.

​
Security audit gate

Every skill passes through an automated audit before installation. The gate blocks skills that contain:
  • Symlinks — symlinks inside a skill archive could escape the skill directory and read arbitrary files
  • Script-like files — executable scripts (*.sh, *.py, *.js, etc.) are not permitted inside skill packages
  • Shell payloads — SKILL.md and manifest files are scanned for embedded shell command patterns
The audit gate is a first line of defence, not a sandbox. Review skill content manually before installing skills from untrusted sources, especially on agents with high autonomy levels or broad tool access.

​
Open skills

By default ZeroClaw only loads skills from paths you explicitly configure. Open skills enables loading from the community skills registry without per-skill path configuration.

​
Enable via config

[skills]
open_skills_enabled = true

​
Enable via environment variable

export ZEROCLAW_OPEN_SKILLS_ENABLED=true
zeroclaw agent
Open skills increases the attack surface. Enable only when you trust the skills registry and have reviewed the audit gate output for each skill in use.

​
Prompt injection mode

When a skill is active, its SKILL.md content is injected into the agent’s system prompt. Two injection modes control how much space the skill content occupies:
ModeDescriptionUse case
fullThe complete SKILL.md is injected verbatimDefault; models with large context windows
compactA summarised or truncated version is injectedLow-context models where token budget is limited
[skills]
injection_mode = "compact"   # full (default) | compact
Choose compact when running against small local models or when multiple skills are active simultaneously and context length is a constraint.

​
Example skill manifest

[skill]
name        = "code-reviewer"
version     = "1.0.0"
description = "Instructs the agent to apply structured code review criteria"
author      = "your-name"
tags        = ["development", "review"]
The corresponding SKILL.md would contain the review criteria, rubric, or checklist the agent should follow when the skill is active.

Build docs developers (and LLMs) love

Get started for freeTalk to us