Using Claude Agent SDK with Marrow MCP

This tutorial shows how to build a small TypeScript app with @anthropic-ai/claude-agent-sdk and connect it to Marrow MCP. The sample app asks Marrow for sales data, lets Claude analyse it, and writes a structured JSON report to ./response/.

What you will build

By the end, you will have a simple agent that:

connects to Marrow MCP from Claude Agent SDK
uses a local MCP server for the app’s helper tools
asks Marrow for export data, then analyses it in the agent
writes the final result as JSON on disk

Prerequisites

Completion of the Marrow MCP Overview setup with an MCP key
Node.js 24 or later
A Marrow organization with an MCP key
An Anthropic API key
yarn (bundled via Corepack - run corepack enable if needed)

Step 1: Clone and install

Start by installing the project:

git clone https://github.com/Deltologic/marrow-mcp-claude-agent-sdk.git
cd marrow-mcp-claude-agent-sdk
yarn install

Step 2: Configure your keys

Copy the example environment file and add your keys:

cp .env.example .env

Set these values in .env:

MARROW_MCP_KEY: your Marrow MCP key
ANTHROPIC_API_KEY: your Anthropic API key

The app reads the Marrow key from the environment and sends it to the hosted MCP server.

Step 3: Understand the Claude entry point

The app logic lives in src/index.ts. Claude Agent SDK handles the agent loop, while Marrow MCP provides the Amazon data. Here is the core shape:

import { createSdkMcpServer, query } from '@anthropic-ai/claude-agent-sdk';

import { systemPrompt } from './prompts/system.prompt.ts';
import { outputTools } from './tools/output.tools.ts';
import { sleepTools } from './tools/sleep.tools.ts';

const localServer = createSdkMcpServer({
    name: 'local-tools',
    tools: [...sleepTools, ...outputTools]
});

const messages = query({
    prompt: 'Analyze the Delto UK account for the last 7 days. Find the top 3 ASINs by total units sold.',
    options: {
        model: 'sonnet',
        systemPrompt,
        mcpServers: {
            marrow: {
                type: 'http',
                url: 'https://mcp.marrow.com/mcp/v1',
                headers: {
                    'marrow-mcp-key': process.env.MARROW_MCP_KEY ?? ''
                }
            },
            'local-tools': localServer
        },
        allowedTools: ['mcp__marrow__*', 'mcp__local-tools__*'],
        maxTurns: 10
    }
});

There are 3 parts to notice:

query(...) starts the Claude run.
mcpServers.marrow connects Claude to the hosted Marrow MCP endpoint.
createSdkMcpServer(...) exposes local helper tools in the same agent session.

The allowedTools list keeps the agent focused, and maxTurns stops the run from continuing forever.

Step 4: Why the system prompt matters

The system prompt in src/prompts/system.prompt.ts tells Claude how to use the tools:

export const systemPrompt = `You are an expert Amazon marketplace data analyst. You have access to Marrow MCP, which provides live Amazon Seller/Vendor data.

Current time: ${new Date().toISOString()}

## Data access pattern

Use \`mcp__marrow__exports_raw_download\` to fetch sales data directly. Analyse the returned data inline - no file download or SQL engine is needed.

## Output rule

After completing your analysis you MUST call \`mcp__local-tools__output_top_asins\` exactly once with all results. Never write the final JSON directly in your response text - always route it through that tool.
`;

This turns a generic agent into a focused workflow:

Marrow MCP is the source of truth for the Amazon data
Claude reasons over the returned data in the conversation
the output tool writes one final JSON artefact to disk

Step 5: Keep the output step explicit

The app uses one local tool for the final handoff: output_top_asins in src/tools/output.tools.ts.

tool(
    'output_top_asins',
    'Log the top-ASINs sales report to console and save it to ./response/<timestamp>.json. Call this exactly once with all results.',
    { asins: z.array(asinSalesSchema) },
    async ({ asins }) => {
        const json = JSON.stringify(asins, null, 2);
        mkdirSync('./response', { recursive: true });
        const filePath = `./response/top-asins-${Date.now()}.json`;
        writeFileSync(filePath, json);
        return { content: [{ type: 'text', text: json }] };
    }
);

The prompt tells Claude to call this tool once, which keeps the final output structured and predictable.

Step 6: Run the app

Start the sample with:

yarn start

The app will:

ask Marrow MCP for the required sales export
let Claude identify the top 3 ASINs by units sold
call output_top_asins once
write the result to ./response/top-asins-<timestamp>.json

A simple mental model

To reuse this pattern in another app, keep the split the same:

Marrow MCP for Amazon data access
Claude Agent SDK for the agent loop and tool orchestration
local tools for file output and any small app-specific helpers

This separation keeps the tutorial simple while still showing a real integration.

Marrow MCP resources

Check the following resources for more information:

MCP server URL: https://mcp.marrow.com/mcp/v1
Interactive Data Scheme
Data Scheme JSON: https://api.marrow.com/api/v1/spec/data-scheme
Need help? Use the contact form

Get Started

Plugins

Trust & Safety

Hologrow Marrow MCP

Using Claude Agent SDK with Marrow MCP

Using Claude Agent SDK with Marrow MCP

What you will build

Prerequisites

Step 1: Clone and install

Step 2: Configure your keys

Step 3: Understand the Claude entry point

Step 4: Why the system prompt matters

Step 5: Keep the output step explicit

Step 6: Run the app

A simple mental model

Marrow MCP resources

​Using Claude Agent SDK with Marrow MCP

​What you will build

​Prerequisites

​Step 1: Clone and install

​Step 2: Configure your keys

​Step 3: Understand the Claude entry point

​Step 4: Why the system prompt matters

​Step 5: Keep the output step explicit

​Step 6: Run the app

​A simple mental model

​Related resources

​Marrow MCP resources

Using Claude Agent SDK with Marrow MCP

What you will build

Prerequisites

Step 1: Clone and install

Step 2: Configure your keys

Step 3: Understand the Claude entry point

Step 4: Why the system prompt matters

Step 5: Keep the output step explicit

Step 6: Run the app

A simple mental model

Related resources

Marrow MCP resources