A proxy network built against AI workloads, not a general-purpose product pivoted to AI

The commercial proxy category was built for a workload shape that doesn't match AI. E-commerce scraping, SEO tracking, and ad-verification workflows drove the vendors through the 2010s. AI data collection showed up in the late 2010s as a minor vertical and was grafted on top of the existing product — residential pools tuned for retail became "AI residential," and datacenter pools designed for bulk scraping became "AI training data" products.

We noticed the mismatch as customers. Single-endpoint per-exit-class routing didn't exist. Metered-residential- plus-unlimited-datacenter pricing that matched the real cost surface didn't exist. Authenticated opt-in SDK sourcing was documented but not always defensible under review. Per-session stickiness for multi-turn agent evaluation was an afterthought.

SquadProxy is an attempt to build the shape that should have existed. The specific product decisions — the five exit classes, the header-based gateway, the plan shape, the 10-country focused instrumentation — all come from "what would an AI engineer actually want?" rather than from what the previous decade of proxy products assumed the customer needed.

SquadProxy Ltd. is the UK entity that contracts with customers, operates the gateway, and maintains the pool-management stack. We're registered at Spaces Shoreditch, 25 Luke Street, London EC2A 4DS, UK. The team sits across London, Berlin, and remote.

The residential pool is operated through an opt-in SDK integration in consumer applications where peers agree to contribute idle-device bandwidth in exchange for disclosed value (SDK features, rewards, ad-free tiers). We describe this arrangement and our provenance posture in detail in the provenance post.

Datacenter, ISP, and mobile pools are contracted through conventional infrastructure providers and carrier partners, with capacity scaled against the 10 markets we focus on.

The gateway is a custom routing fabric that turns one HTTP endpoint plus a few headers into per-request exit-class, country, and session-stickiness selection. The architecture is documented in our API reference.

We support AI training corpus collection, RAG ingestion, LLM evaluation, safety red-team work under documented research programs, benchmark scraping of public academic surfaces, competitive AI intelligence on public surfaces with provider-side rate discipline, and AI agent infrastructure for browser-based and API-based agents.

We decline credential abuse workflows, scraping workflows that violate published Terms of Service, any activity that would constitute unauthorized access under CFAA or equivalent, aggregated personal-data scraping outside documented lawful bases, content categories that are criminally restricted in our home jurisdiction, and any workflow whose actual purpose is extraction of policy-violating model output for downstream sale.

Full details in the Acceptable Use Policy. Safety red-team customers are reviewed at onboarding; see the safety red-team use case for the review framing.