Skip to content
Logo

What is HUSK

HTTP Universal Skill Kernel - turn a folder of agent skills into a working HTTP backend with one process.

HUSK takes a directory of skills and serves every one of them over HTTP. A skill is a folder with a SKILL.md manifest. Its kernel is either a script in any language - bash, Python, TypeScript, anything with a shebang - or an LLM that runs the skill and calls the tools you declare. Run one long-lived Bun process and every skill is a live endpoint. There is no SDK to import and nothing to rewrite: HUSK is the husk that wraps your kernel and publishes it as a service.

bun add -g @elisym/husk
husk init        # creates ./skills with a starter skill
husk serve       # every skill is now an HTTP endpoint on :3000
curl -X POST http://localhost:3000/skills/uppercase --data 'hello'
# HELLO

Three ideas

  1. A skill is a folder, not a plugin. The format is compatible with Agent Skills (Claude Code, Codex, Cursor): one SKILL.md manifest, any language. Existing skill folders serve unchanged.
  2. A script, or an LLM with tools. A skill is a script you wrote, or mode: llm - where an LLM runs the skill using the SKILL.md body as its system prompt and calls the tools you declare. Same folder, same HTTP surface.
  3. One long-lived node on Bun. Start the server and every skill is reachable over HTTP. The same folder also runs as a one-shot serverless function or inside a container with no redesign.

What HUSK is not

  • Not a framework you code against - there is no SDK, no base class, no decorator. A skill is data plus a kernel.
  • Not a model host - for mode: llm skills HUSK calls your chosen provider with your API key; it bundles no model and no provider SDK.
  • Not a marketplace - HUSK is the HTTP shell. Discovery, payments, and identity live in separate protocols, not here.

Pick your path

Reading this as an automated agent? The docs expose an MCP source - use the "Ask AI" button, or fetch /llms-full.txt for the full text.

Edit on GitHub