Proxy infrastructure for AI agents

Give your agents
a real network.

Residential, ISP, mobile and datacenter egress behind a single endpoint. Target a country, pin a session across a multi-step task, and cap spend — without writing proxy plumbing into every tool you ship.

Start with Starter — $49/mo Read the docs

Card required· Credentials provisioned after payment· HTTP, HTTPS & SOCKS5

agent — egress trace

# tool call 1 — fetch pricing page $ curl -x $PROXY_URL https://example.com/pricing 200 via 172.98.14.203 · us · residential # tool call 2 — same session, new request $ curl -x $PROXY_URL https://example.com/api/plans 200 sticky · session=a41f · 6s old # tool call 3 — geo moved to de $ curl -x $PROXY_URL -H 'X-AIP-Country: de' ... 403 target blocked by origin · rotated ip 200 retry 1 · 89.22.7.14 · de

p50 420 ms countries 195 IPs 40M+

Plays nice with

Playwright Puppeteer Crawl4AI LangChain OpenAI Agents SDK LlamaIndex MCP clients

The problem

Browsers get blocked. Agents get expensive.

An agent that retries blindly burns budget and context. The proxy layer is usually the part that decides whether an autonomous workflow finishes — or loops until it times out.

One IP, one ban

Ten parallel tool calls from a single datacenter IP look exactly like a scraper. Rotate egress per request, or keep a stable identity when the flow needs it.

Geo-gated data

Pricing, inventory and SERPs differ by country and city. Target egress at the request level instead of standing up a fleet of regional runners.

Runaway loops

A broken retry loop is a billing incident. Set hard request and bandwidth ceilings per key, and let the proxy return a typed error before the agent spends it.

Products

Every egress type. One credential.

Switch product per request with a username flag. Same gateway, same auth, same error contract.

Residential rotating

Real consumer devices for sites that block everything else.

40M+ IPs across 195 countries
Rotate per request or pin for 30 min
City and ASN targeting
HTTP/SOCKS5

$4.50/ GB

Pay as you go. Volume tiers from 50 GB.

ISP static

Datacenter speed with a residential-grade reputation.

Dedicated IP, no rotation
Unmetered bandwidth
US, UK, DE, FR, NL, CA, AU
Sub-100 ms p50

$1.80/ IP / mo

Billed monthly. Cancel anytime.

Mobile 4G / 5G

Carrier-grade NAT for the hardest targets.

4G and 5G carrier IPs
Rotate per request
High trust score
HTTP/SOCKS5

$8.00/ GB

Best for social and app-store data.

Datacenter budget

Cheap, fast, predictable egress for friendly targets.

Shared and dedicated pools
Unmetered bandwidth
20+ regions
Lowest p50 latency

$0.60/ GB

Start at $19/mo for a shared pool.

Built for tool calls

The details that make a loop finish.

Everything below is designed around machine callers: deterministic strings, typed failures, and limits that fail closed.

One endpoint, flag-driven

Encode product, country, city and session in the username. No SDK, no side-channel config, no per-language proxy code.

user-country-de-session-ab12-product-res

Sticky sessions

Pin egress to one IP for up to 30 minutes so a login, a cart and a checkout all come from the same place.

session=a41f · 30 min TTL

Hard spend caps

Per-key request and GB ceilings. Once the cap is hit the gateway returns 402 immediately instead of silently billing.

cap → 402 quota_exceeded

Typed errors

blocked, auth_failed, geo_unavailable, quota_exceeded. Your agent can branch on a code instead of parsing HTML.

X-AIP-Error

Tool schemas included

Drop-in JSON schema for function calling, plus a documented MCP transport so any client can request egress as a tool.

function calling ready

Usage you can meter

Per-key request and bandwidth counters exposed as JSON, so cost per task can be compared with task success rate.

GET /v1/usage

Five-minute integration

Point the SDK you already use.

Nothing to install and no proxy SDK. A proxy URL is the whole integration, so it works with any HTTP client, browser automation stack or LLM tool runner.

Pick a plan and pay — credentials provision on payment
Set HTTP_PROXY / HTTPS_PROXY and run your agent
Add a country or session flag when a task needs it
Cap spend per key before the first production run

# credentials from your dashboard
export U=user-country-de
export P=your-proxy-password

curl -x "http://$U:$P@gate.aiproxies.dev:8000" \
  https://httpbin.org/ip

# pin one IP for the whole task
U=user-country-us-session-ab12
curl -x "http://$U:$P@gate.aiproxies.dev:8000" \
  -H "X-AIP-Product: isp" https://httpbin.org/ip

import httpx

U = "user-country-gb-session-ab12"
P = "your-proxy-password"
proxy = f"http://{U}:{P}@gate.aiproxies.dev:8000"

with httpx.Client(proxy=proxy, timeout=30) as client:
    r = client.get("https://example.com/pricing")
    r.raise_for_status()
    # branch on the typed error, not on the HTML
    if r.headers.get("X-AIP-Error") == "blocked":
        retry_with_new_session()

import { ProxyAgent } from "undici";

const agent = new ProxyAgent(
  "http://" + process.env.PROXY_USER +
  ":" + process.env.PROXY_PASS +
  "@gate.aiproxies.dev:8000"
);

const res = await fetch("https://example.com/api/plans", {
  dispatcher: agent
});

console.log(res.status, res.headers.get("x-aip-ip"));

import { chromium } from "playwright";

const browser = await chromium.launch({
  proxy: {
    server: "http://gate.aiproxies.dev:8000",
    username: "user-country-de-session-ab12",
    password: process.env.PROXY_PASSWORD
  }
});

const page = await browser.newPage();
await page.goto("https://example.com");

Network

Coverage where agents get blocked.

Pool sizes and latency below are placeholders in this repo until measured from the live pool — see the README before publishing numbers.

40M+residential IPs

195countries

2,400+ASNs

420 msp50 residential latency

Per-request targeting

Country, region, city and ASN flags. Unavailable combinations fail fast with geo_unavailable rather than hanging until timeout.

Session pools

Unlimited named sessions, 30-minute pinning, sticky per key by default. Sessions are released on completion or TTL expiry.

Health and rotation

Unhealthy peers are drained from rotation automatically. Blocked requests are retried once on a fresh IP before returning an error.

Pricing

Pick a plan. Scale when the loop works.

Every plan is prepaid monthly. Your proxy credentials are provisioned as soon as the first payment clears — there is no free tier and no free trial.

Starter

$49/mo

For a single agent in production.

10 GB residential included
5 dedicated ISP IPs
20 concurrent sessions
All countries
Email support

Get started

Scale popular

$299/mo

For fleets of agents and parallel tasks.

100 GB residential included
50 dedicated ISP IPs
200 concurrent sessions
Mobile pool access
Usage API + webhooks
Priority support

Get started

Enterprise

Custom

For committed volume and compliance reviews.

Volume pricing
Dedicated account team
SOC 2 report on request
Private gateway / VPC peering
Custom AUP review

Talk to sales

Billed monthly in USD, excluding tax, cancel anytime. Unused ISP IPs can be swapped once per month. Overage billed at list rate.

FAQ

Questions worth answering up front.

Is there a free tier or a trial?

No. Every residential, ISP and mobile IP we allocate costs us money, so all plans are paid and there is no free tier or free trial.

Signing up creates an account and an API key, but your proxy credentials are only provisioned once the first payment clears. If you need to evaluate the network before committing, talk to sales about a paid pilot.

Do you log the requests my agent makes?

No request content is logged on residential or mobile traffic. We keep connection metadata — timestamp, key, exit IP, bytes, target hostname — for 24 hours, strictly for abuse handling and billing reconciliation, then delete it.

ISP and datacenter traffic keeps the same 24-hour window. Enterprise plans can extend retention for their own audit purposes, never ours.

What is acceptable use?

Public web data collection, QA and monitoring, ad verification, price and inventory intelligence, and geo testing are all fine. Credential stuffing, carding, scraping authenticated accounts you do not own, bypassing access controls, and anything targeting minors are not.

The acceptable use policy is the binding version. We act on abuse reports within one business day.

How does a session stay sticky?

Add -session-<id> to the username and every request with that id exits from the same IP until the 30-minute TTL expires or the session is released. Without a session flag, the gateway rotates per request.

Can I use this with an LLM function-calling loop?

Yes — that is the primary use case. The docs include a ready JSON tool schema and an egress_fetch example that returns status, final URL, body and the exit IP, so the model gets a deterministic result instead of a proxy string to reason about.

What happens when I hit my spend cap?

The gateway responds 402 with X-AIP-Error: quota_exceeded. No silent overage, no surprise invoice. Raise the cap from the dashboard and the same key resumes immediately.

Do you offer SOCKS5?

Yes, on port 1080 for all products. HTTP and HTTPS proxies are on 8000. Both accept the same username flags and password.

Is there a refund policy?

Unused monthly plans are refundable within 14 days, pro-rated. Pay-as-you-go bandwidth is refundable for failed requests that were billed — open a ticket with the request ids and we credit them.

Your agent is one payment away.

Choose a plan and check out. Your gateway credentials are provisioned as soon as payment clears — usually within a minute.

Start with Starter Read the docs

Give your agentsa real network.