LangChain

Give a LangChain agent web search and clean Markdown page fetches through the Web Access API.

The langchain-string package wraps the Web Access API as two LangChain tools. StringSearch runs a search and returns the results as JSON; StringFetch fetches a URL and returns the page as Markdown, with proxy rotation, anti-bot handling, CAPTCHA solving and JavaScript rendering happening server-side.

Both are ordinary BaseTool implementations with sync and async paths, so they work in any LangChain agent.

Install

pip install -U langchain-string

Set your API key, or pass api_key=SecretStr("...") to either tool:

export STRING_API_KEY=...

Search

from langchain_string import StringSearch

search = StringSearch(max_results=5)
print(search.invoke({"query": "who makes the best web scraping api"}))

The tool returns the search response as JSON: results holds the ranked pages, and a Google search can also carry the knowledge panel, local listings, AI overviews and other surfaces the results page showed around them. A query answered by a panel alone therefore still has an answer, even though results is empty.

The agent can set query, engine, country and language per call.

Field	Default	Description
`query`	—	Required. The search query. Quoted phrases and `site:` pass through to the engine.
`engine`	`google`	`google`, `duckduckgo`, `brave` or `mojeek`. Only Google returns the surfaces above.
`country`	`US`	ISO 3166-1 alpha-2 country code used to localize the results.
`language`	—	Language tag such as `en` or `pt-br`.
`max_results`	—	Constructor only. Organic results wanted, 1 to 50. See below.

max_results is a billing decision, which is why the agent cannot set it

Without it, a search returns one results page, about ten results, billed as one search. With it, Google is paged until it has that many, and each page is billed as one search.

Fetch

from langchain_string import StringFetch

fetch = StringFetch(main_content_only=True)
print(fetch.invoke({"url": "https://example.com"}))

The agent can set url, and execute_js when a JavaScript-rendered page comes back empty, and country_code to route the request through a specific country. Everything else is set when you construct the tool.

Field	Default	Description
`url`	—	Required. The `http`/`https` URL to fetch.
`execute_js`	`false`	Render the page in a browser before capturing it.
`country_code`	—	ISO 3166-1 alpha-2 country to route the request through.
`markdown_mode`	`full`	Constructor only. `full` or `readable`. See Markdown for LLMs.
`main_content_only`	`false`	Constructor only. Strip navigation, footers and other page chrome.
`max_content_length`	—	Constructor only. Truncate the page to this many characters.

When the site answers with an error status, the output says so above the page it returned, so the agent can tell a 404 page from an article.

Both tools in an agent

StringWebAccessToolkit returns both tools and passes its configuration to each.

from langchain.agents import create_agent
from langchain_string import StringWebAccessToolkit

agent = create_agent(
    model="anthropic:claude-sonnet-5",
    tools=StringWebAccessToolkit(max_results=5).get_tools(),
    system_prompt=(
        "You research questions using the live web. Search first, then read the pages "
        "worth reading, and cite every URL you used."
    ),
)

result = agent.invoke(
    {"messages": [{"role": "user", "content": "What does String charge for a CAPTCHA solve?"}]}
)
print(result["messages"][-1].content)

Failures reach the model

API and network failures are returned to the model as the tool's output, carrying the reason the API gave: the rejected field, the status the destination answered with, or an account problem such as an insufficient balance. The model can then change its call instead of the run ending. Pass handle_tool_error=False to raise instead.

A missing API key is not treated this way. It raises before any request, because no retry fixes it.

Shared options

Field	Default	Description
`api_key`	`STRING_API_KEY`	Held as a `SecretStr`, so it stays out of reprs and traces.
`base_url`	`https://request.usestring.ai/v1`	Change only to point at a different String deployment.
`timeout`	`600.0`	Seconds. A browser-rendered fetch of a slow page can take minutes.
`http_client` / `http_async_client`	—	Your own `httpx` clients, for proxies or custom transports.