Claude Agent SDK

Give an agent you are building the news index, in TypeScript or Python.

Setup

1. Install the SDK for your language.
2. Declare the server in the options you pass to the query/agent call.
3. Let the model choose the tools, or restrict it with an allow-list if you want a narrower agent.

import { query } from "@anthropic-ai/claude-agent-sdk";

const result = query({
  prompt: "What happened with semiconductor export controls this week? Cite outlets.",
  options: {
    mcpServers: {
      unzoi: {
        type: "http",
        url: "https://api.unzoi.com/mcp",
        headers: { "x-api-key": process.env.UNZOI_KEY! },
      },
    },
    // Optional: a research agent rarely needs to check its own billing.
    allowedTools: [
      "mcp__unzoi__search_news",
      "mcp__unzoi__list_stories",
      "mcp__unzoi__get_article",
      "mcp__unzoi__find_related",
    ],
  },
});

for await (const message of result) console.log(message);

Python

          import os, anyio
from claude_agent_sdk import query, ClaudeAgentOptions

async def main():
    options = ClaudeAgentOptions(
        mcp_servers={
            "unzoi": {
                "type": "http",
                "url": "https://api.unzoi.com/mcp",
                "headers": {"x-api-key": os.environ["UNZOI_KEY"]},
            }
        },
    )
    async for message in query(
        prompt="Summarise this week's coverage of port congestion, by region.",
        options=options,
    ):
        print(message)

anyio.run(main)
        

What differs here

• Tool names are namespaced by the SDK as mcp__<server>__<tool>, which is what an allow-list has to match — the bare tool name will not.
• An allow-list is worth setting. account_status and top_headlines are rarely what a research agent needs, and every tool you leave in is a tool the model can spend quota on.
• The server instructions string is surfaced to the model automatically, so it already knows what this corpus is without you restating it in the prompt.

Next

• Get a key at console.unzoi.com, if you have not.
• The tool reference — what the model is reading when it chooses between the 18 tools.
• Stories for context windows — the one habit that most changes how well an agent works against this API.
• Limits — an agent that loops can spend a month's credits faster than you expect.
• Claude Agent SDK's own MCP documentation — authoritative for anything about the client itself, including config paths, which move.

Endpoint, for copying: https://api.unzoi.com/mcp