LangChain

/** * LangChain with Boxline: two LangChain tools backed by Boxline, given to a LangChain agent (createAgent) on your own * OpenAI key: `boxline_fetch` reads a page in a sandboxed browser, and `boxline_browser_agent` hands a whole browser * task (clicking, typing, several pages) to a Boxline agent run. * *   npm install && npx tsx index.ts        (BOXLINE_API_KEY and OPENAI_API_KEY in the environment; LC_MODEL, QUESTION) * * Writes output/result.json: the answer and every tool call. */import { mkdirSync, writeFileSync } from "node:fs";import { join } from "node:path";import { Boxline } from "@boxline/sdk";import { ChatOpenAI } from "@langchain/openai";import { createAgent, tool } from "langchain";import { z } from "zod";const out = process.env.OUTPUT_DIR ?? "output";const question = process.env.QUESTION ?? 'On books.toscrape.com, what does "Sapiens: A Brief History of Humankind" cost, and how many are in stock?';const bx = new Boxline();const boxlineRuns: { id: string }[] = [];const boxlineFetch = tool(  async ({ url }) => {    const page = await bx.fetch(url, { format: "markdown", links: true });    return JSON.stringify({ url: page.finalUrl, status: page.status, title: page.title, content: page.content.slice(0, 12_000), links: (page.links ?? []).slice(0, 80) });  },  {    name: "boxline_fetch",    description: "Open a web page in a real browser (in a Boxline sandbox) and return it as Markdown with its links.",    schema: z.object({ url: z.string().describe("an http(s) address") }),  },);const boxlineBrowserAgent = tool(  async ({ task }) => {    const run = await bx.agent.run({ task, maxSteps: 20 });    boxlineRuns.push({ id: run.id });    const done = await bx.agent.wait(run.id, { timeoutMs: 8 * 60_000 });    return done.status === "completed" ? (done.result ?? "") : `The browser agent ${done.status}: ${done.error ?? ""}`;  },  {    name: "boxline_browser_agent",    description: "Give a browser task that needs clicking, typing or several pages to an AI agent with its own browser. Returns its answer.",    schema: z.object({ task: z.string().describe("what to do, with the site's address") }),  },);const agent = createAgent({  // The Responses API: newer OpenAI models take function tools only there.  model: new ChatOpenAI({ model: process.env.LC_MODEL ?? "gpt-6-luna", useResponsesApi: true }),  tools: [boxlineFetch, boxlineBrowserAgent],  systemPrompt: "Answer from pages you read with the tools. Prefer boxline_fetch; use boxline_browser_agent only when a task needs clicking or typing.",});const result = await agent.invoke({ messages: [{ role: "user", content: question }] });const calls = result.messages.flatMap((m: any) => (m.tool_calls ?? []).map((c: any) => ({ tool: c.name, input: c.args })));const answer = result.messages.at(-1)?.text ?? ""; // .text joins the message's text blocksfor (const c of calls) console.log(`→ ${c.tool} ${JSON.stringify(c.input)}`);console.log(`\n${answer}`);mkdirSync(out, { recursive: true });writeFileSync(join(out, "result.json"), JSON.stringify({ question, answer, calls, agentRuns: boxlineRuns }, null, 2));

"""LangChain with Boxline: two LangChain tools backed by Boxline, given to a LangChain agent (create_agent) on yourown OpenAI key: `boxline_fetch` reads a page in a sandboxed browser, and `boxline_browser_agent` hands a whole browsertask (clicking, typing, several pages) to a Boxline agent run.    pip install -r requirements.txt && python main.py     (BOXLINE_API_KEY and OPENAI_API_KEY; LC_MODEL, QUESTION)Writes output/result.json: the answer and every tool call."""import jsonimport osfrom pathlib import Pathfrom langchain.agents import create_agentfrom langchain.tools import toolfrom langchain_openai import ChatOpenAIfrom boxline import Boxlineout = Path(os.environ.get("OUTPUT_DIR", "output"))question = os.environ.get("QUESTION", 'On books.toscrape.com, what does "Sapiens: A Brief History of Humankind" cost, and how many are in stock?')bx = Boxline()boxline_runs = []@tooldef boxline_fetch(url: str) -> str:    """Open a web page in a real browser (in a Boxline sandbox) and return it as Markdown with its links."""    page = bx.fetch(url, format="markdown", links=True)    return json.dumps({"url": page["finalUrl"], "status": page["status"], "title": page["title"], "content": page["content"][:12_000],                       "links": (page.get("links") or [])[:80]})@tooldef boxline_browser_agent(task: str) -> str:    """Give a browser task that needs clicking, typing or several pages to an AI agent with its own browser. Returns its answer."""    run = bx.agent.run(task, max_steps=20)    boxline_runs.append({"id": run["id"]})    done = bx.agent.wait(run["id"], timeout=480)    if done["status"] != "completed":        return f"The browser agent {done['status']}: {done.get('error') or ''}"    return done["result"] or ""agent = create_agent(    # The Responses API: newer OpenAI models take function tools only there.    ChatOpenAI(model=os.environ.get("LC_MODEL", "gpt-6-luna"), use_responses_api=True),    tools=[boxline_fetch, boxline_browser_agent],    system_prompt="Answer from pages you read with the tools. Prefer boxline_fetch; use boxline_browser_agent only when a task needs clicking or typing.",)result = agent.invoke({"messages": [{"role": "user", "content": question}]})calls = [{"tool": c["name"], "input": c["args"]} for m in result["messages"] for c in (getattr(m, "tool_calls", None) or [])]answer = result["messages"][-1].text  # joins the message's text blocksfor c in calls:    print(f"-> {c['tool']} {json.dumps(c['input'])}")print(f"\n{answer}")out.mkdir(parents=True, exist_ok=True)(out / "result.json").write_text(json.dumps({"question": question, "answer": answer, "calls": calls, "agentRuns": boxline_runs}, indent=2, ensure_ascii=False))

# Node 22 or newerexport BOXLINE_API_KEY=…        # your Boxline API keyexport OPENAI_API_KEY=…         # your own OpenAI keynpm install @boxline/sdk @langchain/core@1.2.13 @langchain/openai@1.6.0 langchain@1.5.14 tsx@4.23.15 zod@4.4.3npx tsx index.ts

# Python 3.10 or newerexport BOXLINE_API_KEY=…        # your Boxline API keyexport OPENAI_API_KEY=…         # your own OpenAI keypip install boxline-sdk langchain==1.4.3 langchain-openai==1.6.6python main.py