Proxies for Claude and the Anthropic API

Why proxy Claude / Anthropic API traffic

1. Regional policy eval. Claude's content policy and retrieval (where applicable) apply regional variation. Residentials anchored to target countries give the authentic origin for measuring the delta.

2. Multi-turn agent benchmarks. Claude's agentic capabilities (Computer Use, tool use) run sessions that benefit from IP stability. ISP static keeps multi-turn sessions coherent.

3. Safety red-team work. Anthropic coordinates closely with external safety evaluators. Red-team workflows running jurisdictional guardrail mapping anchor on residentials per target country.

4. Eval-fleet concurrency. Continuous eval pipelines that run thousands of prompts across tens of countries need concurrency that single-account rate limits alone can't support without careful distribution.

What this page isn't

• Not a way to access Claude in restricted countries.
• Not for circumventing Anthropic's Acceptable Use Policy.
• Not a rate-limit bypass for a single API key.

Recommended configuration

import httpx

PROXY = "http://USER:PASS@gateway.squadproxy.com:7777"

def eval_claude(prompt: str, country: str, model: str = "claude-opus-4-7"):
    return httpx.post(
        "https://api.anthropic.com/v1/messages",
        json={
            "model": model,
            "max_tokens": 1024,
            "messages": [{"role": "user", "content": prompt}],
        },
        headers={
            "x-api-key": ANTHROPIC_API_KEY,
            "anthropic-version": "2023-06-01",
            "X-Squad-Class": "residential",
            "X-Squad-Country": country,
            "X-Squad-Session": "per-request",
        },
        proxies=PROXY,
        timeout=120,
    ).json()

For multi-turn workflows (Computer Use, agent benchmarks), swap to X-Squad-Class: isp and X-Squad-Session: sticky-30m.

Claude-specific eval considerations

• Anthropic's safety-first positioning means policy variation across regions is a specifically reportable finding.
• Computer Use — evaluating Claude Computer Use in realistic browsing conditions benefits from residential routing; the target sites the agent visits classify cloud IPs differently from residential.
• Tool use with external APIs — the tool calls made by Claude agents route through your proxy layer; the origin appears to the tool targets as whichever exit class you configured.

Plans that fit

See pricing. Team plan is the typical starting point for a continuous Claude eval pipeline; Lab plan covers safety red-team coordination at frontier-lab scale.

Related

• Safety red-team use case
• LLM evaluation use case
• ChatGPT API landing
• Residential vs datacenter routing