kernl
guides·

Building a coding agent in < 10 min

dremnik

We just published a new guide: Charger — A Coding Agent.

It walks through building a complete coding agent that can read, write, and execute code in a sandboxed environment using Daytona. The whole thing takes about 10 minutes.

What you'll build

  • An agent with file system, shell, and git tools running in isolated Daytona sandboxes
  • Web search via MCP (using Parallel AI)
  • A custom todo toolkit for task tracking
  • A streaming API server with Hono
  • A chat UI using the Vercel AI SDK

Key patterns

The guide covers some patterns we think are important:

Lazy sandbox provisioning — sandboxes are created on first tool use, not upfront. The sandbox ID is stored in thread context, so it survives across HTTP requests.

Credentials through context — git tokens flow through context server-side, never touching the LLM. The model just calls git_clone(url, path) and the tool injects auth. No accidental leakage in responses or logs.

Type-safe context — toolkit contexts are intersected (SandboxContext & TodoContext), so TypeScript ensures your tools have access to the state they need.

Try it

pnpm create kernl@latest charger
cd charger
npm install -g @kernl-sdk/cli
kernl add toolkit daytona
kernl add toolkit parallel

Then follow the full guide or check out the complete example.