Epi

The registry covers LLM fine-tuning (LoRA, DPO, GRPO), distributed training (FSDP, DeepSpeed), vector databases (Qdrant, Pinecone, FAISS, Chroma), financial modeling (DCF, LBO, comps, 3-statement), OSINT and web pentesting, Docker management, GitHub PR lifecycles, and creative generation via ComfyUI, Blender, and Stable Diffusion.

ML engineers running training and inference pipelines, investment analysts building Excel valuation models, security researchers doing authorized pentests and supply-chain forensics, developers delegating coding to Claude Code or Codex, and creators producing video, music, and diagrams through Hermes Agent.

Skills run inside Hermes Agent on Linux, macOS, or Windows. Most load automatically by trigger phrase; some require CLI prerequisites (python3, curl, node, ffmpeg) or environment variables such as SHOPIFY_ACCESS_TOKEN, NOTION_API_KEY, or a 1Password service account token, collected via interactive setup prompts.

Nearly all skills carry MIT licenses, with Anthropic-adapted finance skills under Apache-2.0. One exception is the powerpoint skill, which is proprietary with terms in LICENSE.txt. Underlying services like Pinecone, Lambda Labs, or Stripe may charge their own usage fees.

Training skills require PyTorch, transformers, and often CUDA GPUs — torchtitan scales 8 to 512+ GPUs, while Unsloth and QLoRA target limited VRAM. Vector search skills need clients like qdrant-client or chromadb. Several skills are pure Python stdlib, including osint-investigation and domain-intel.