Real-World Pain Points

Real-World Pain Points

Project: https://github.com/Narwhal-Lab/MagicSkills

For many teams building multi-agent systems, the first thing that gets out of control is not the model, but skill management.

The same skill directory often needs to serve agent applications such as Codex, Cursor, and Claude Code, while also being used by multiple agents created inside frameworks such as LangChain and LangGraph.

What usually follows is not reuse, but duplication: the same skill directory gets copied into multiple Agent projects, and once that skill directory needs to change, you have to maintain it in multiple places. It quickly forks.

How MagicSkills Solves This

MagicSkills is not trying to be yet another Agent framework. It adds a local-first skill infrastructure layer for multi-Agent projects.

In one sentence: build a skill once, reuse it across every Agent.

More specifically:

• MagicSkills first aggregates installed skills into one shared skill pool
• It then creates dedicated Skills collections for different Agents, exposing only the capabilities each Agent actually needs
• Finally, based on how each runtime integrates skills, it either syncs to AGENTS.md / CLAUDE.md or exposes them as tool / function

One lower-level implementation detail is worth mentioning: MagicSkills aggregates installed skills into a unified `Allskills` view, but for external communication, “shared skill pool” is the easier concept to understand.

An Extreme Example

Assume you want to do something as extreme as possible, and also as representative as possible of MagicSkills' value:

you want one single skill directory, with no copying at all, to serve the following Agent applications and the agents built inside Agent frameworks:

• Codex
• Cursor
• Claude Code
• Windsurf
• Aider
• AutoGen
• CrewAI
• LangChain
• LangGraph
• Haystack
• Semantic Kernel
• smolagents
• LlamaIndex

In that case, the whole process can be broken down into four steps.

1. Install MagicSkills

git clone https://github.com/Narwhal-Lab/MagicSkills.git
cd MagicSkills
pip install -e .

2. Install the Required Skills into the Shared Skill Pool

# Option 1: install skills from a local directory
magicskills install skill_template

# Option 2: install skills from GitHub
magicskills install anthropics/skills

Notes:

• The first command installs skills from the local skill_template, such as c_2_ast
• The second command installs more skills from GitHub, such as pdf, docx, brand-guidelines, doc-coauthoring, and canvas-design
• The install path can be specified with -t, or you can use the default path
• No matter where they are installed, MagicSkills aggregates those installed skills into one shared skill pool

3. Create Dedicated `Skills` Collections for Each Agent

This is the most important part of the design.

Instead of copying a skill directory into some Agent-specific folder, you first create a named `Skills` collection for each Agent and se