langchain-ai-skills-framework

Overview

• langchain-ai-skills-framework loads Agent Skills from SKILL.md files and serves them via LangChain tools.
• Skills come from two sources:
  • Plugin marketplace (MarketplaceDirectoryLoader) — shared skills loaded from filesystem or GitHub, organized as plugins/<plugin>/skills/<skill>/SKILL.md.
  • User-persisted skills (MongoPluginSkillLoader) — stored in MongoDB across three collections (plugin_skills, plugin_references, plugin_scripts).
• CompositeSkillLoader merges both sources with precedence: user → shared DB → marketplace.
• All skill tools are scoped by plugin_name.

Skill authoring

• See docs/skill-authoring.md for required frontmatter, naming rules, and examples.

GitHub authentication for remote plugins

• When PLUGINS_MARKETPLACE uses github://..., provide a token via SKILLS_GITHUB_TOKEN (preferred) or GITHUB_TOKEN (fallback).
• Expected GitHub directory format: github://<owner>/<repo>/<path>?ref=<branch> (for example: github://my-org/ai-plugin-marketplace/plugins?ref=main).
• Supported token types: fine-grained Personal Access Token (PAT) and GitHub App installation token.
• Recommended usage:
  • Local development: fine-grained PAT scoped to the required repositories.
  • CI (GitHub Actions): workflow GITHUB_TOKEN when permissions are sufficient.
  • Long-running services: short-lived GitHub App installation tokens.

Quick start

• make init – initialize the local dev environment.
• make up – start the dev container.
• make run-pre-commit – run lint/type/security suite.
• make tests – run dockerized pytest.
• make build – build sdist/wheel packages.