What is skill-retriever?

Skill Retriever is a plugin for Hermes Agent, an AI agent framework, that helps the agent pick which skills to load for a given user request. Skills here are packaged bundles of instructions the agent can use, similar in idea to plugins, and a system with hundreds of them needs a fast way to figure out which handful actually apply to the task at hand. The README explains that an older approach walked a tree of skill categories using repeated calls to a language model, which could take roughly 243 calls per query and was too slow for real use. Skill Retriever instead builds a single flat index of all available skills once, ahead of time, then narrows that list down to the top fifty likely candidates, and finally makes just one language model call to pick the best three to twenty skills for the current query. Each selected skill comes back with a load priority marked as must, should, or consider, a confidence level of high, medium, or low, and a short reason explaining why it was chosen. That information gets written into the user's message as a hint block the agent can act on, and skills are then loaded with a call to skill_view. You install it with pip, and Skill Retriever registers itself as a Hermes plugin automatically without any extra plugin development. It hooks into several points in Hermes: a pre_llm_call hook that injects a curated bundle on every turn, a subagent dispatch hook for per-task bundles, and a planning phase hook for splitting skill coverage across phases of a longer plan. It also logs every skill selection to a JSONL file for later review. Skill Retriever needs Hermes Agent version 0.18 or newer, Python 3.10 or newer, roughly 50 megabytes for its flat index, and access to an OpenAI-compatible language model endpoint such as LongCat or OpenAI itself. The project's own code is released under the MIT license, though it notes that community-contributed skills bundled separately may carry their own licenses.