# Quickstart

Get a skill serving over HTTP in under a minute. HUSK runs on [Bun](https://bun.sh); install it first if you have not.

## Install

```bash
bun add -g @elisym/husk
```

Check it:

```bash
husk --help
```

## Create a project

```bash
husk init my-agent
cd my-agent
```

This writes a `skills/` folder with a starter `hello` skill:

:::file-tree

* +my-agent
  * +skills
    * +hello the starter skill
      * SKILL.md the manifest
      * run.sh the kernel

:::

## Serve it

```bash
husk serve
```

```
  HUSK serving 1 skill(s)
  from /…/my-agent/skills

  POST   /skills/hello              hello

  → http://localhost:3000
```

In another terminal:

```bash
curl -X POST http://localhost:3000/skills/hello --data 'world'
# you said: world
```

Open `http://localhost:3000` for the index, or `http://localhost:3000/openapi.json` for the generated spec.

## Add your own skill

Scaffold one in the language you like:

```bash
husk new weather --lang python
```

Edit `skills/weather/run.py` - read input from stdin, print the result to stdout - and reload:

```bash
husk serve --watch
```

:::tip
`--watch` reloads skills when files change, so you can edit a kernel and re-`curl` without restarting the server.
:::

## Test without HTTP

`husk call` invokes a skill exactly the way the server would, but from the command line - handy for development and for serverless entry points:

```bash
echo "hello" | husk call uppercase -i -
husk call weather -i "Berlin"
```

## Start a prompt

Prefer to let your AI assistant drive? Hand it this:

```txt
Read https://docs.husk.systems and scaffold a HUSK skill in Python that takes a city name on stdin and prints the current weather as JSON, then serve it with `husk serve` and show me a working curl.
```

## Next steps

* [Anatomy of a skill](/skills/anatomy) - the folder, the manifest, the script.
* [The kernel I/O contract](/skills/kernel) - exactly how input and output flow.
* [One skill, three runtimes](/serve/runtimes) - ship the same folder as a function or container.