Agents & Skills
Create custom agents and discover reusable skills for Claude Code.
Agents
Agents are autonomous configurations with specialized system prompts, tool permissions, and memory scopes.
Browsing Agents
The Agents page shows all agents grouped by scope (user, project, plugin) with stat cards. Each agent card displays the name, description, scope badge, and configured tools.
Creating an Agent
Click "New Agent" to open the wizard:
- Name & Description — agent identity
- Scope — user (global) or project (version-controlled)
- Tools — list of allowed tools and disallowed tools
- Model — Claude model to use
- Content — system prompt (markdown)
Agent Configuration
Agents support YAML frontmatter in markdown files:
yaml
---
description: Code review specialist
tools: Read, Grep, Glob
disallowed-tools: Write, Bash
permission-mode: default
model: claude-sonnet-5
skills: code-review, testing
memory: user
---
Review code for quality, security, and maintainability.| Property | Description |
|---|---|
tools | Allowed tools list |
disallowed-tools | Blocked tools list |
permission-mode | Permission level for the agent |
model | Claude model |
skills | Skills the agent can use |
hooks | Before/after hooks (command or prompt type) |
memory | Memory scope (none, user, session) |
Storage
| Location | Scope |
|---|---|
~/.claude/agents/ | User agents |
.claude/agents/ | Project agents |
Skills
Skills are reusable knowledge and workflow modules — markdown files with metadata that extend Claude's capabilities.
Browsing Skills
The Skills page has two tabs:
- Installed — skills on your system with dependency status badges
- Discover — browse and install from the skills.sh registry
Each skill card shows the name, description, scope badge, and tags (context, tools, fork, hidden, user-invocable).
Skill Metadata
Skills use YAML frontmatter:
yaml
---
name: code-review
description: Review code for quality issues
version: 1.0.0
allowed-tools: Read, Grep, Glob
user-invocable: true
argument-hint: "path to review"
---| Property | Description |
|---|---|
allowed-tools | Tools the skill can access |
user-invocable | Whether users can call it directly as /skill-name |
argument-hint | Hint text for skill arguments |
context | Fork context for the skill |
disable-model-invocation | Prevent Claude from auto-invoking |
Storage
| Location | Scope |
|---|---|
~/.claude/skills/ | User skills |
.claude/skills/ | Project skills |
Tips
- Plugin agents and skills are read-only — view but not edit.
- Skills show dependency status (green = satisfied, amber = missing) to help troubleshoot.
- Agents can reference skills in their frontmatter to compose capabilities.