# Claude Agent SDK source verification

- **Agent pattern:** In-process MCP
- **Search surfaces:** Web and news
- **Saved artifact:** `claude-source-verification.md`


# What you'll build

A Claude Agent SDK workflow that turns a broad public-policy claim into a source-verification memo. SerpApi tools are exposed as an in-process SDK MCP server, so the agent can search without running a separate service.

**Default brief:** Test whether right-to-repair rules are rapidly converging across the United States. Separate confirmed facts, overstatements, and jurisdictional differences.


# How the agent works

1.  **Mount the tools.** The recipe registers typed web and news searches on a local `serpapi` MCP server.
2.  **Split the evidence.** Web search finds primary or durable sources; news search checks recent developments.
3.  **Reconcile the claim.** Claude exposes conflicts and jurisdictional gaps before returning a final memo within sixteen turns.


# Core agent setup

This excerpt shows how the SerpApi tools become an in-process SDK MCP server. The complete script also validates credentials, collects the final SDK result, and writes the memo.

``` python
serpapi_server = create_sdk_mcp_server(
    name="serpapi",
    version="1.0.0",
    tools=[
        web_search(provider="claude-agent-sdk"),
        news_search(provider="claude-agent-sdk"),
    ],
)
options = ClaudeAgentOptions(
    model=MODEL,
    system_prompt="Search before concluding and expose conflicting evidence.",
    mcp_servers={"serpapi": serpapi_server},
    allowed_tools=["mcp__serpapi__web_search", "mcp__serpapi__news_search"],
    max_turns=16,
)

async for message in query(prompt=PROMPT, options=options):
    ...
```

[View the complete `main.py` →](https://github.com/serpapi/serpapi-search-tools-python/blob/main/cookbook/claude-agent-sdk/main.py)


# Run the recipe

Set `SERPAPI_API_KEY` and `ANTHROPIC_API_KEY` in the repository-root `.env`, then run:

``` bash
uv run --isolated --no-project --with 'serpapi-search-tools[claude-agent-sdk]' --with python-dotenv cookbook/claude-agent-sdk/main.py
```

Optional controls: `ANTHROPIC_MODEL`, `COOKBOOK_PROMPT`, and `COOKBOOK_OUTPUT_DIR`.

**No separate MCP process:** The SDK creates the MCP server in-process and allows only the two SerpApi tools declared by the recipe.


# Inspect the result

The recipe writes `cookbook-output/claude-source-verification.md`. Expect confirmed facts, overstatements, jurisdiction differences, conflicting evidence, and direct source URLs.

The recipe saves the final result emitted by the SDK and also prints it to the terminal.


# SerpApi tools used

[web_search](../../reference/web_search.md#serpapi_search_tools.web_search) finds primary and durable policy sources, while [news_search](../../reference/news_search.md#serpapi_search_tools.news_search) tracks recent right-to-repair developments. Both tools are exposed through the in-process MCP server, helping the agent compare established rules with current reporting before it classifies the claim.

**Framework pattern:** The in-process MCP setup follows Anthropic's custom-tools example and provides SerpApi web and news tools to the agent.

- [Open the runnable recipe](https://github.com/serpapi/serpapi-search-tools-python/tree/main/cookbook/claude-agent-sdk)
- [View Anthropic's source example](https://github.com/anthropics/claude-agent-sdk-python#custom-tools-as-in-process-sdk-mcp-servers)