SkillGarden (Python)

Part of LoopEngine — a runtime for defining and running AI agents through a transparent ReAct loop — handling SKILL.md-based skill discovery and lazy loading. Works standalone too, no dependency on LoopEngine itself.

Ported line-for-line from the TypeScript implementation. See the root README for the pitch, the prior-art landscape, and the scope decisions — this file only covers what's specific to running the Python code.

Install

pip install -e ".[dev]"

Quickstart

PYTHONPATH=src python3 examples/quickstart.py

from skillgarden import SkillGarden

garden = SkillGarden(dirs=["examples/skills"])
result = garden.build_index()

print(garden.invoke("greet", "world"))
# -> "Hello, world!"

eligible = garden.eligible_for(result.included, ["apps/web/index.tsx"])
# -> includes "deploy:web", since its `paths` glob matches

CLI

skillgarden add firecrawl --agent file-agent
# -> Added file-agent:firecrawl -> skills/file-agent/firecrawl/SKILL.md

Copies a skill from the bundled skillgarden/registry/ into skills/<agent>/<skill>/, ready for discover_skill_files to pick up as <agent>:<skill>. Pass --dir <path> to use a skills root other than ./skills, and --force to overwrite an existing install.

Test

pytest

Status

Loader, budget-aware index, directory discovery with colon-namespacing, and path-conditional activation are real and tested. Published as skillgarden on PyPI.