Residential proxies for AI agents and AI data pipelines

"AI proxy" means three different things, and only one of them is this page. First, an IP proxy that AI agents and data pipelines use to reach websites through other addresses and locations, so a browser agent in a US workflow arrives from a US consumer connection rather than a cloud datacenter. That is what TrueProxies sells and what the rest of this page covers.

Second, an LLM API gateway or reverse proxy such as LiteLLM, Kong or APISIX, which sits between your code and a model provider to route, log and meter model calls. It moves prompts, not page loads, and it does not change the IP a website sees. Third, a "proxy URL" in AI roleplay chat apps, which is a relay for a chat API key. Neither of those is an IP proxy, and a residential proxy does nothing for them.

Match the proxy to how the workload moves bytes. Agents that think between page loads are latency-bound and light on traffic; crawlers that feed a training set are the opposite. The product that fits is the one whose billing unit matches that shape.

MB per page equals GB per 1,000 pages, so one measured number sets the budget. We loaded the same 50 public pages (news, documentation, product pages, blog posts and forum threads, one per domain) three ways through the TrueProxies residential gateway on October 1, 2026: a plain HTTP fetch of the HTML, a headless Chromium with images, fonts and media blocked, and a full Chromium render. Bytes are what the gateway billed, scaled from the browser's own network accounting by the measured factor of 1.381.

A full render moved 31.7x the traffic of the HTML fetch. Blocking images, fonts and media cut the full-render figure by 59%, which is the single cheapest optimisation for a browser agent that only needs text. The scripts, the URL list and the raw results are public, under the MIT licence.

Concurrent connections aren't billed separately. Plans allow up to 3,000 connections (Datacenter IPv6: 3,000). The cap is per service, not per account, and it is an abuse limit rather than a performance tier: a fleet that needs more runs a second service.

An HTTP worker holds about one connection at a time. A browser is greedier: over a 10-page subset routed through a counting relay, Chromium peaked at 22 concurrent tunnels on one page load with a full render and 26 with assets blocked, with medians of 12 and 6. The peak is set by how many hosts a page pulls from at once, not by what you block, and Chromium's own ceiling is 32 sockets per proxy. Plan on roughly 115 browser agents per service at the measured peak of 26 tunnels, more if your targets are lighter.

Size the fleet on the Free hour before buying: one hour of Residential IPv4 Unlimited at 10 Mbps with up to 100 connections. Then buy the Traffic pack whose per-GB price matches the volume you measured.

A multi-step task (log in, search, open a result, read it) should keep one exit IP from start to finish. A mid-task IP change looks like an account takeover to the site and resets whatever state the agent built. Give every agent its own session ID, set the lifetime to at least the longest task, and rotate between tasks rather than inside them.

Sticky sessions on GB-based plans last from 1 minute to 24 hours and are set on the proxy username: `USER-country-de-session-agent01-lifetime-600` keeps one German exit for ten minutes under the session ID `agent01`. Set the lifetime explicitly rather than relying on a default. For independent fetches (RAG retrieval, a crawl of unrelated URLs) leave the session off and let the gateway rotate per request.

Datacenter IPv6 gives a crawler private /48 in your chosen city, rotating by default with sticky sessions on demand, from $3.99/day. It only works where the target serves IPv6: a host with no AAAA record cannot be reached from an IPv6-only exit at all. Run the target through the IPv6 test first, and keep residential IPv4 for everything that fails it.

Each integration page below was run through the TrueProxies gateway with the version named on the page: the config, the measured traffic per page and the errors we hit, with fixes. They are written for self-hosted and open-source setups; the Firecrawl and Browser Use cloud plans bundle their own proxies.

Collect public data only. Respect robots.txt, site terms and text-and-data-mining opt-outs, keep request rates reasonable, and never use an agent to access accounts, credentials or paywalled content you are not entitled to. A proxy changes the network location of a request; it does not change what you are allowed to do with the response. The acceptable use policy applies to every service.