Skilljack MCP

An MCP server that jacks Agent Skills directly into your LLM's brain.

Recommended: For best results, use an MCP client that supports tools/listChanged notifications (e.g., Claude Code). This enables dynamic skill discovery - when skills are added or modified, the client automatically refreshes its understanding of available skills.

Features

• Dynamic Skill Discovery - Watches skill directories and automatically refreshes when skills change

• Tool List Changed Notifications - Sends tools/listChanged so clients can refresh available skills

• Skill Tool - Load full skill content on demand (progressive disclosure)

• MCP Resources - Access skills via skill:// URIs with batch collection support

• Resource Subscriptions - Real-time file watching with notifications/resources/updated

Motivation

This repo demonstrates a way to approach integrating skills using existing MCP primitives.

MCP already has the building blocks:

• Tools for on-demand skill loading (the skill tool with dynamically updated descriptions)

• Resources for explicit skill access (skill:// URIs)

• Notifications for real-time updates (tools/listChanged, resources/updated)

• Prompts for explicitly invoking skills by name (/my-server-skill)

This approach provides separation of concerns. Rather than every MCP server needing to embed skill handling, the server acts as a dedicated 'skill gateway'. Server authors can bundle skills alongside their MCP servers without modifying the servers themselves. If MCP registries support robust tool discovery, skill tools become discoverable like any other tool.

Installation

Or run directly with npx:

From Source

Usage

Configure one or more skills directories containing your Agent Skills:

Each directory is scanned along with its .claude/skills/ and skills/ subdirectories for skills. Duplicate skill names are handled by keeping the first occurrence.

Windows note: Use forward slashes in paths when using with MCP Inspector:

How It Works

The server implements the Agent Skills progressive disclosure pattern with dynamic updates:

1. At startup: Discovers skills from configured directories and starts file watchers

2. On connection: Skill tool description includes available skills metadata

3. On file change: Re-discovers skills, updates tool description, sends tools/listChanged

4. On tool call: Agent calls skill tool to load full SKILL.md content

5. As needed: Agent calls skill-resource to load additional files

Tools vs Resources

This server exposes skills via both tools and resources:

• Tools (skill, skill-resource) - For your agent to use autonomously. The LLM sees available skills in the tool description and calls them as needed.

• Resources (skill:// URIs) - For manual selection in apps that support it (e.g., Claude Desktop's resource picker). Useful when you want to explicitly attach a skill to the conversation.

Most users will rely on tools for automatic skill activation. Resources provide an alternative for manual control.

Tools

skill

Load and activate an Agent Skill by name. Returns the full SKILL.md content.

Input:

Output: Full SKILL.md content including frontmatter and instructions.

skill-resource

Read files within a skill's directory (scripts/, references/, assets/, snippets/, etc.).

This follows the Agent Skills spec's progressive disclosure pattern - resources are loaded only when needed.

Read a single file:

Read all files in a directory:

Returns all files in the directory as multiple content items.

List available files (pass empty path):

Security: Path traversal is prevented - only files within the skill directory can be accessed.

Resources

Skills are also accessible via MCP Resources using skill:// URIs.

URI Patterns

Resource Subscriptions

Clients can subscribe to resources for real-time updates when files change.

Capability: resources: { subscribe: true, listChanged: true }

Subscribe to a resource:

Receive notifications when files change:

Unsubscribe:

How it works:

1. Client subscribes to a skill:// URI

2. Server resolves URI to file path(s) and starts watching with chokidar

3. When files change, server debounces (100ms) and sends notification

4. Client can re-read the resource to get updated content

Security

Skills are treated as trusted content. This server reads and serves skill files directly to clients without sanitization. Only configure skills directories containing content you trust.

Protections in place:

• Path traversal prevention (symlink-aware)

• File size limits (1MB default, configurable via MAX_FILE_SIZE_MB env var)

• Directory depth limits

• Skill content is confined to configured directories

Not protected against:

• Malicious content within trusted skill directories

• Prompt injection via skill instructions (skills can influence LLM behavior by design)

Dynamic Skill Discovery

The server watches skill directories for changes. When SKILL.md files are added, modified, or removed:

1. Skills are re-discovered from all configured directories

2. The skill tool's description is updated with current skill names and metadata

3. tools/listChanged notification is sent to connected clients

4. Clients that support this notification will refresh tool definitions

Skill Metadata Format

The skill tool description includes metadata for all available skills in XML format:

This metadata is dynamically updated when skills change - clients supporting tools/listChanged will automatically refresh.

Skill Discovery

Skills are discovered at startup from the configured directories. For each directory, the server checks:

• The directory itself for skill subdirectories

• .claude/skills/ subdirectory

• skills/ subdirectory

Each skill subdirectory must contain a SKILL.md file with YAML frontmatter including name and description fields.

Testing

Related

• Agent Skills Specification

• MCP TypeScript SDK

• Example MCP Clients