How to give AI agents and LLM tools a proxy

• An active CleanProxies plan: static ISP proxies for fixed US IPs, or rotating residential / mobile proxies for pools. The trial works too.
• Your credential line from the dashboard's proxy manager: host (or IP), port, username and password.
• Your agent framework and the language it runs in (examples use Python).
• A list of the tools that fetch web pages, so you know which ones need the proxy.

New to proxies? For the overview, read Using proxies with cURL, Python and Node.js and Rotating vs sticky sessions. This guide goes deeper on AI agents and LLM tools only.

Most tools want the same details as a URL: http://USERNAME:PASSWORD@HOST:PORT for HTTP(S) or socks5://USERNAME:PASSWORD@HOST:PORT for SOCKS5. The examples on this page use those placeholders; copy the host, port, username and password from your CleanProxies dashboard. To convert a whole list at once, paste it into the proxy list formatter.

1. 1

   Find the web-facing layer

   Browser tools usually run Playwright or Chromium; fetch tools usually run requests, httpx, aiohttp or Node fetch.
2. 2

   Configure the proxy there

   Use the library's own proxy option (see the Playwright and Requests guides), or environment variables.
3. 3

   Keep LLM calls direct

   If you use environment variables, add the LLM provider's API host to NO_PROXY so model traffic doesn't use your proxy bandwidth.
4. 4

   Verify from inside the agent

   Ask the agent to fetch https://api.ipify.org with its own tools and report the result.

Setting HTTP_PROXY/HTTPS_PROXY is the quickest way to proxy tools you don't control, but not every client reads them:

export HTTP_PROXY="http://USERNAME:PASSWORD@HOST:PORT"
export HTTPS_PROXY="http://USERNAME:PASSWORD@HOST:PORT"
# add your LLM provider's API host, plus local services
export NO_PROXY="localhost,127.0.0.1,api.openai.com,api.anthropic.com"

• Python requests and HTTPX: yes, by default (HTTPX via trust_env=True).
• aiohttp: only when the session is created with trust_env=True.
• Node.js fetch and http: only with NODE_USE_ENV_PROXY=1 (Node 24+ and 22.21+); axios reads them.
• Go's default transport and .NET HttpClient: yes. Java: no, it uses https.proxyHost-style system properties.
• Chromium-based browsers launched by agents: no; set the browser's proxy option instead.

Most browser agents launch Playwright or Chromium under the hood. Where the framework exposes launch or context options, pass Playwright's proxy object; where it accepts an existing browser or context, create it yourself with the proxy and hand it over.

from playwright.async_api import async_playwright

PROXY = {"server": "http://HOST:PORT", "username": "USERNAME", "password": "PASSWORD"}

async def make_context():
    p = await async_playwright().start()
    browser = await p.chromium.launch()
    return await browser.new_context(proxy=PROXY)  # pass this to the agent's browser tool

Chromium can't authenticate to SOCKS5 proxies, so use the HTTP form of your credentials for any browser-based agent. For browser-use specifically, see the browser-use guide.

LangChain's requests tools (TextRequestsWrapper and the RequestsGetTool family) call Python requests synchronously, so they pick up HTTP_PROXY/HTTPS_PROXY. The async path uses aiohttp, which ignores those variables; pass an aiosession created with a proxy or trust_env=True.

WebBaseLoader has a proxies argument for its requests session.

from langchain_community.document_loaders import WebBaseLoader

proxy = "http://USERNAME:PASSWORD@HOST:PORT"
loader = WebBaseLoader(
    "https://example.com/pricing",
    proxies={"http": proxy, "https": proxy},
)
docs = loader.load()

import aiohttp
from langchain_community.utilities.requests import TextRequestsWrapper

session = aiohttp.ClientSession(trust_env=True)  # reads HTTPS_PROXY / NO_PROXY
wrapper = TextRequestsWrapper(aiosession=session)

• One identity per agent: a dedicated ISP IP (or sticky residential line) per agent instance, never shared between agents running in parallel.
• Location by task: generate residential lines for the country a task needs, from 190+ countries and a 50M+ IP pool.
• Fetch-only tools that read many public pages can use rotating residential lines; browser tools that click through flows should not.
• Log the exit IP at the start of each run so a misconfigured tool shows up immediately.

Ask the agent to call its web tool on https://api.ipify.org and report the result, and check that your LLM requests still go direct (they should keep working with the proxy turned off).

Open What is my IP? through the proxy to see the address and headers a website receives, or paste up to ten lines into the free proxy checker to test connectivity, latency and exit IP before you start.

Still stuck? Our 24/7 support team answers fast: contact us with the error and the line you used (leave the password out).