langchain-agentskills

agentskills.io support for LangChain and LangGraph.

Install

pip install langchain-agentskills

Quick Start

Toolkit (classic)

from langchain.agents import create_agent
from langchain_agentskills import SkillsToolkit

toolkit = SkillsToolkit(directories=["./skills"])

agent = create_agent(
    model="gpt-5.2",
    tools=toolkit.get_tools(),
    system_prompt=toolkit.get_instructions(),
)

Middleware (recommended)

from langchain.agents import create_agent
from langchain_agentskills import DirectorySkillLoader, SkillMiddleware

loader = DirectorySkillLoader("./skills")

agent = create_agent(
    model="gpt-5.2",
    middleware=[SkillMiddleware(loader=loader)],
)

The middleware automatically injects skill descriptions into the system prompt and registers load_skill, read_skill_resource, and run_skill_script tools.

Skill directory structure

skills/
  my-skill/
    SKILL.md          # Frontmatter + instructions
    references/       # Resource files the agent can read
      api-spec.yaml
    scripts/          # Executable scripts the agent can run
      deploy.sh

SKILL.md uses YAML frontmatter:

---
name: my-skill
description: Short description for the agent
version: "1.0"
tags: [deployment, ci]
---

# Instructions

Detailed instructions for the agent...

Custom loaders

Implement SkillLoader to load skills from any source (Git, S3, database, etc.):

from langchain_agentskills import SkillLoader

class MyGitLoader(SkillLoader):
    def list_skills(self): ...
    def load_skill(self, name): ...
    def read_resource(self, skill_name, resource_name): ...
    def read_script(self, skill_name, script_name): ...
    def has_skill(self, name): ...

Use with the toolkit or middleware:

# Toolkit
toolkit = SkillsToolkit(loaders=[MyGitLoader(...)])

# Middleware
middleware = SkillMiddleware(loader=MyGitLoader(...))

License

MIT