From Zero to One: Build a Complete Clean Agent Project

This tutorial walks through a full clean-SDK project, not just a quick demo. We stay on createAgentSdk() from start to finish.

What you will build

By the end, you will have a terminal-based agent that:

1. loads a local JSON config
2. keeps multi-turn session history
3. streams responses
4. calls local tools
5. uses clean skills
6. can be pointed at a specific workspace
7. can later grow into MCP, memory, swarm, and computer-use workflows

1. Install dependencies

npm install actoviq-agent-sdk zod

2. Prepare a config file

Create agent.settings.json:

{
  "HADAMARD_BASE_URL": "https://your-model-endpoint.example.com/v1",
  "HADAMARD_AUTH_TOKEN": "your-token",
  "HADAMARD_MODEL": "your-model-name"
}

The nested env form also works.

3. Build the first runnable app

Create src/app.ts:

import path from 'node:path';
import process from 'node:process';
import readline from 'node:readline/promises';
import { stdin as input, stdout as output } from 'node:process';

import { z } from 'zod';
import {
  createAgentSdk,
  createHadamardFileTools,
  loadJsonConfigFile,
  skill,
  tool,
} from 'actoviq-agent-sdk';

const CONFIG_PATH = path.resolve(process.cwd(), 'agent.settings.json');
const WORK_DIR = process.cwd();

await loadJsonConfigFile(CONFIG_PATH);

const addNumbers = tool(
  {
    name: 'add_numbers',
    description: 'Add two numbers.',
    inputSchema: z.object({
      a: z.number(),
      b: z.number(),
    }),
  },
  async ({ a, b }) => ({ sum: a + b }),
);

const releaseCheck = skill({
  name: 'release-check',
  description: 'Check release-readiness concerns.',
  prompt: 'You are running the release-check skill.\n\nTask:\n$ARGUMENTS',
  inheritDefaultTools: false,
  inheritDefaultMcpServers: false,
  allowedTools: [],
});

const sdk = await createAgentSdk({
  workDir: WORK_DIR,
  tools: [...createHadamardFileTools({ cwd: WORK_DIR }), addNumbers],
  skills: [releaseCheck],
});

const rl = readline.createInterface({ input, output });
const session = await sdk.createSession({ title: 'My Clean Agent' });

console.log('Agent started. Type exit to quit.');
console.log(`session.id = ${session.id}`);

try {
  while (true) {
    const prompt = (await rl.question('\nYou> ')).trim();
    if (!prompt) continue;
    if (['exit', 'quit', '/exit', ':q'].includes(prompt.toLowerCase())) break;

    if (prompt.startsWith('/skill ')) {
      const args = prompt.slice('/skill '.length).trim();
      const result = await session.runSkill(
        'release-check',
        args || 'Summarize the release checks for this project.',
      );
      console.log(`\nAgent> ${result.text}`);
      continue;
    }

    const stream = session.stream(prompt, {
      systemPrompt: 'You are a clear, reliable, concise engineering assistant.',
    });

    output.write('\nAgent> ');
    for await (const event of stream) {
      if (event.type === 'response.text.delta') {
        output.write(event.delta);
      }
    }

    const result = await stream.result;
    output.write('\n');

    if (result.toolCalls.length > 0) {
      console.log('Tools used:', result.toolCalls.map(call => call.name));
    }
  }
} finally {
  rl.close();
  await sdk.close();
}

4. Run it

npx tsx src/app.ts

Or add a script:

{
  "scripts": {
    "dev": "tsx src/app.ts"
  }
}

Then run:

npm run dev

5. Add the next layers

After the first version works, expand in this order:

1. more local tools
2. more clean skills
3. session recovery with sdk.sessions.list() and sdk.resumeSession(...)
4. memory and session.extractMemory()
5. swarm teammates
6. workspace isolation

6. Where to look next

Useful repository examples:

1. examples/hadamard-quickstart.ts
2. examples/hadamard-skills.ts
3. examples/hadamard-agent-helpers.ts
4. examples/hadamard-swarm.ts