Getting Started

Install the CLI

npm install -g @kernl-sdk/cli

Create a project

kernl init my-project
cd my-project

Your first agent

Open src/agents/jarvis.ts. This is your agent:

import { Agent } from "kernl";
import { anthropic } from "@kernl-sdk/ai/anthropic";

import { math } from "@/toolkits/math";

export const jarvis = new Agent({
  id: "jarvis",
  name: "Jarvis",
  instructions: "You are Jarvis, a helpful assistant.",
  model: anthropic("claude-sonnet-4-5"),
  toolkits: [math],
});

The starter project includes a simple math toolkit—now let's add something more useful.

Add toolkits

Toolkits give your agent capabilities. You can grab pre-built ones from the registry or write your own.

From the marketplace

Browse kernl.sh/marketplace and add with a single command:

kernl add toolkit github

This downloads the toolkit to toolkits/, installs dependencies, and shows any environment variables to configure.

Your own tools

Create a tool with tool() and group them into a Toolkit:

import { tool, Toolkit } from "kernl";
import { z } from "zod";

const greet = tool({
  id: "greet",
  description: "Greet someone by name",
  parameters: z.object({ name: z.string() }),
  execute: async (ctx, { name }) => `Hello, ${name}!`,
});

export const hello = new Toolkit({
  id: "hello",
  description: "Greeting tools",
  tools: [greet],
});

Then wire toolkits into your agent:

import { github } from "@/toolkits/github";
import { hello } from "@/toolkits/hello";

export const jarvis = new Agent({
  id: "jarvis",
  name: "Jarvis",
  instructions: "You are Jarvis, a helpful assistant.",
  model: anthropic("claude-sonnet-4-5"),
  toolkits: [github, hello],
});

Run your agent

There are two ways to run an agent:

Blocking — wait for the full response:

const result = await jarvis.run("What can you help me with?");
console.log(result.response);

Streaming — receive events as they happen:

const stream = jarvis.stream("Tell me a joke");

for await (const event of stream) {
  if (event.kind === "text.delta") {
    process.stdout.write(event.text);
  }
}

The starter template runs this in src/index.ts. Try it:

pnpm start

Project structure

The CLI scaffolds a sensible basic project structure for you:

src/
├── agents/
│   └── jarvis.ts
├── toolkits/
│   ├── math/
│   │   ├── index.ts
│   │   ├── add.ts
│   │   └── ...
│   └── github/        ← added via registry
│       └── index.ts
└── index.ts

For real applications, you'll likely want to run your agent inside a server. Check out the microprojects for production-ready examples:

Jarvis — Chief of Staff for dev teams with Linear + GitHub toolkits
Watson — Customer discovery agent with memory and workflows

Both follow the same pattern: a server/ running Hono + kernl, and a web/ frontend chatbot that hits the agent API.

Next steps

Agents — deeper dive into agent configuration
Toolkits — create custom toolkits or connect to MCP servers
Memory — give your agent persistent memory
Threads — manage execution state