Build your own agent with the Agent SDK

The steps

One step at a time. When a step’s check matches, tick it and the next one opens.

1. 01

   Step 1 of 4: Install the SDK in the kit

   One npm package, with Claude Code bundled.

   Unzip the kit and open a terminal in its folder. The package bundles a native Claude Code binary, so most setups need nothing else.

   TerminalCopy

   – What happens: In the horizonlux-claude-code-day-51 folder:

   Run: $ npm install @anthropic-ai/claude-agent-sdk

   Check it worked

   package.json now lists @anthropic-ai/claude-agent-sdk under dependencies.

   It worked — next step Not yet
2. 02

   Step 2 of 4: Set your API key

   In the same shell that will run the agent.

   The SDK reads the key from the environment of the process that runs it. It doesn’t load .env files, and the key never goes in a committed file.

   For macOS, Linux and WSL

   TerminalCopy

   Run: $ export ANTHROPIC_API_KEY=your-api-key

   For Windows

   PowerShellCopy

   Run: PS> $env:ANTHROPIC_API_KEY = "your-api-key"

   Check it worked

   Works in every shellCopy

   node -e "console.log(Boolean(process.env.ANTHROPIC_API_KEY))"

   It prints true without showing the key.

   It worked — next step Not yet Back
3. 03

   Step 3 of 4: Read the agent

   A prompt, two options and a loop.

   agent/fix-bugs.mjs · JavaScriptCopy

   // Day 51: a small agent built with the Claude Agent SDK.
   //
   // It follows the official quickstart
   // (https://code.claude.com/docs/en/agent-sdk/quickstart), written as plain
   // JavaScript so `node` runs it without a TypeScript step.
   //
   // Run it from the project folder, after `npm install @anthropic-ai/claude-agent-sdk`:
   //   node agent/fix-bugs.mjs
   // ANTHROPIC_API_KEY must be set in the same shell. The SDK does not read .env files.
   import { query } from '@anthropic-ai/claude-agent-sdk';
   
   // The agent loop: query() streams messages while Claude works.
   for await (const message of query({
     prompt: 'Review agent/practice.js for bugs that would cause crashes. Fix any issues you find.',
     options: {
       allowedTools: ['Read', 'Edit', 'Glob'], // pre-approved: read, edit and find files
       permissionMode: 'acceptEdits', // approve file edits without asking
     },
   })) {
     // Print only the readable parts: Claude's text and the tools it calls.
     if (message.type === 'assistant' && message.message?.content) {
       for (const block of message.message.content) {
         if ('text' in block) {
           console.log(block.text);
         } else if ('name' in block) {
           console.log(`Tool: ${block.name}`);
         }
       }
     } else if (message.type === 'result') {
       console.log(`Done: ${message.subtype}`);
     }
   }
   

   query starts the agent loop and streams messages as Claude works. allowedTools pre-approves reading, finding and editing files, and acceptEdits approves edits without asking. The docs’ quickstart uses TypeScript; this is the same code as plain JavaScript.

   In order:

   1. You

      You run the script

   2. Claude

      query() starts the loop

   3. Tool

      Read, Glob and Edit

      Pre-approved, no prompts

   4. Shell

      A result message

      The script prints Done: and the result

   Check it worked

   You can point to the prompt, the tool list and the line that prints Done:.

   It worked — next step Not yet Back
4. 04

   Step 4 of 4: Run it and check the fix

   The agent edits practice.js, then you check it.

   agent/practice.js has two functions that crash: one on an empty list, one when the client is missing.

   TerminalCopy

   Run: $ node agent/fix-bugs.mjs

   – What happens: Claude’s reasoning and each tool it calls print as it works.

   Output: Done: success

   Check it worked

   Both functions in agent/practice.js now handle the input that crashed them, and npm test still passes.

   If it didn’t work

   It fails with Not logged in or Invalid API key.

   Set ANTHROPIC_API_KEY in the shell that runs the agent, then run it again.

   It worked — finish Not yet Back