Claude Agent SDK - CoreWeave Docs

The Claude Agent SDK enables you to quickly build agent applications powered by Claude. You can integrate Weave into you Claude agents to automatically trace their calls, including agent queries, model responses, tool use, and multi-turn conversations. Weave displays the captured data in the Agents view of your project.

Trace Claude Agent SDK agents with Weave

Python
TypeScript

The Weave SDK autopatches the Claude Agent SDK for Python, letting you capture traces from your Claude agents with minimal setup.This guide shows how to initialize Weave and run a multi-turn Claude agent with MCP tools through query(). Weave automatically traces the conversation, model calls, and tool calls end-to-end.

Prerequisites

A CoreWeave Forge account and API key set as a WANDB_API_KEY environment variable.
An Anthropic API key set as an ANTHROPIC_API_KEY environment variable.
Python 3.10+.

Weave integrates with the @anthropic-ai/claude-agent-sdk to automatically trace query() calls, including agent spans, model responses, and tool calls.

Prerequisites

A CoreWeave Forge account and API key set as a WANDB_API_KEY environment variable.
An Anthropic API key set as an ANTHROPIC_API_KEY environment variable.
Node.js 18+.
@anthropic-ai/claude-agent-sdk version 0.3.178 or later.

Install packages

Install the following packages in your developer environment. The weave package captures traces, claude-agent-sdk provides the agent runtime, and the remaining packages support the example tool.

pip install weave claude-agent-sdk requests

npm install weave @anthropic-ai/claude-agent-sdk zod

Initialize Weave in your code

Python
TypeScript

Add weave.init to the project, update your CoreWeave Forge team and project names, and then build an agent the way you normally would. weave.init enables the autopatching that captures traces from the Claude Agent SDK.This example defines a wikipedia_search MCP tool and runs a three-turn conversation. Each turn is a separate query() call, but later turns pass resume with the first turn’s session ID so that all turns group as one conversation in the Weave Agents view. The first two turns trigger Wikipedia lookups, and the third uses the previous conversation context to produce a summary without a tool call.

import anyio
import requests
import weave

from claude_agent_sdk import (
    ClaudeAgentOptions,
    ResultMessage,
    create_sdk_mcp_server,
    query,
    tool,
)

weave.init("<your-team>/<your-project-name>")

@tool(
    "wikipedia_search",
    "Search Wikipedia for a topic and return its title and intro paragraph.",
    {"query": str},
)
async def wikipedia_search(args: dict) -> dict:
    r = requests.get(
        "https://en.wikipedia.org/w/api.php",
        params={
            "action": "query", "generator": "search", "gsrsearch": args["query"], "gsrlimit": 1,
            "prop": "extracts", "exintro": True, "explaintext": True, "format": "json",
        },
        headers={"User-Agent": "weave-demo"},
    ).json()
    page = next(iter(r["query"]["pages"].values()))
    return {"content": [{"type": "text", "text": f"{page['title']}: {page['extract']}"}]}

wiki_server = create_sdk_mcp_server(
    name="wiki",
    version="1.0.0",
    tools=[wikipedia_search],
)

async def main():
    session_id = None

    for question in [
        "Who founded Anthropic?",
        "What is Claude (the AI assistant)?",
        "Summarize what we discussed in one sentence.",
    ]:
        options = ClaudeAgentOptions(
            mcp_servers={"wiki": wiki_server},
            allowed_tools=["mcp__wiki__wikipedia_search"],
            resume=session_id,
        )
        print(f"USER: {question}")

        async for message in query(prompt=question, options=options):
            if isinstance(message, ResultMessage):
                session_id = message.session_id
                print(f"AGENT: {message.result}\n")


anyio.run(main)

Each query() call produces an invoke_agent root span. Because the follow-up turns resume the same session, Weave stamps the same gen_ai.conversation.id on all spans and groups them as one conversation in the Agents view. ResultMessage.session_id carries the ID to reuse, and resume accepts None on the first turn, which starts a new session.Without resume, each turn starts its own session and appears as a separate conversation. ClaudeSDKClient is the exception: it holds one session open for the lifetime of the async with block, so turns you send through the same client are already grouped. Weave patches both paths.

Weave automatically instruments query() via module loader hooks. The required setup differs slightly between module systems. For more information on CommonJS versus ESM and how Weave’s loader hooks work, see the TypeScript SDK third-party integration guide.

CommonJS projects: No extra configuration needed. Require weave before @anthropic-ai/claude-agent-sdk so auto-instrumentation runs first.
ESM projects: Start Node with the --import=weave/instrument flag so the instrumentation loads before any other modules.

This example defines a wikipedia_search MCP tool and runs a three-turn conversation. Each turn is a separate query() call, but later turns pass resume with the first turn’s session_id so that all turns group as one session in the Weave Agents view. The first two turns trigger Wikipedia lookups, and the third uses the previous conversation context to produce a summary without a tool call.

main.mjs

import * as weave from "weave";
import { createSdkMcpServer, query, tool, type Options } from "@anthropic-ai/claude-agent-sdk";
import { z } from "zod";

const wikipediaSearch = tool(
    "wikipedia_search",
    "Search Wikipedia for a topic and return its title and intro paragraph.",
    { query: z.string().describe("The topic to search for") },
    async ({ query }) => {
    const url = new URL("https://en.wikipedia.org/w/api.php");
    url.search = new URLSearchParams({
        action: "query",
        generator: "search",
        gsrsearch: query,
        gsrlimit: "1",
        prop: "extracts",
        exintro: "true",
        explaintext: "true",
        format: "json",
    }).toString();

    const response = await fetch(url, { headers: { "User-Agent": "weave-demo" } });
    const data = await response.json();
    const page = Object.values(data.query.pages)[0] as { title: string; extract: string };
    return { content: [{ type: "text" as const, text: `${page.title}: ${page.extract}` }] };
    },
);

const wikiServer = createSdkMcpServer({
    name: "wiki",
    version: "1.0.0",
    tools: [wikipediaSearch],
});

async function main() {
    await weave.init("<your-team>/<your-project-name>");

    const baseOptions: Options = {
    model: "claude-sonnet-4-5",
    maxTurns: 4,
    mcpServers: { wiki: wikiServer },
    allowedTools: ["mcp__wiki__wikipedia_search"],
    };

    const questions = [
    "Who founded Anthropic?",
    "What is Claude (the AI assistant)?",
    "Summarize what we discussed in one sentence.",
    ];

    let sessionId: string | undefined;

    for (const prompt of questions) {
    const options: Options = sessionId
        ? { ...baseOptions, resume: sessionId }
        : baseOptions;

    console.log(`USER: ${prompt}`);
    for await (const message of query({ prompt, options })) {
        if (message.type === "system" && message.subtype === "init") {
        sessionId ??= message.session_id;
        }
        if (message.type === "result" && message.subtype === "success") {
        console.log(`AGENT: ${message.result}\n`);
        }
    }
    }
}

main().catch(console.error);

Each query() call produces an invoke_agent root span. Because the follow-up turns resume the same session_id, Weave stamps the same gen_ai.conversation.id on all spans and groups them as one session in the Agents view.Save the example as main.mjs and run it with the --import=weave/instrument flag so the loader hook runs before any other module:

node --import=weave/instrument main.mjs

See your agent traces in the Agents view

After the script runs, weave.init() prints a link to your project. Open the Agents view to inspect:

A session containing the conversation’s turns.
Each turn rendered as an invoke_agent span with nested chat and execute_tool children.
The full input, model, output, token usage, and tool results at each step.

For details about viewing Agents data in Weave, see View agent activity.

​Trace Claude Agent SDK agents with Weave

​Prerequisites

​Prerequisites

​Install packages

​Initialize Weave in your code

​See your agent traces in the Agents view

Trace Claude Agent SDK agents with Weave

Prerequisites

Prerequisites

Install packages

Initialize Weave in your code

See your agent traces in the Agents view