tierproxy

tierproxy — Python SDK

🚧 Preview release. Gateway is not yet generally available. Join the waitlist at [email protected]. SDK is functional but tierproxy doctor against a live gateway requires invitation.

Premium multi-provider proxy infrastructure for AI/ML pipelines. Built for engineers who measure cost, latency, and success rate twice — and write Python.

Install

pip install tierproxy

Quickstart — five-second flavor

import tierproxy
r = tierproxy.get("https://example.com", country="US")
print(r.text)

That's it. (Set TIERPROXY_API_KEY env var first.)

Three lines, persistent session

from tierproxy import TierProxy
with TierProxy() as g:
    print(g.me.get().client_id)
    r = g.get("https://example.com", country="US", session_id="s1")

Auto-pick the cheapest healthy upstream every request

g = TierProxy(routing="cheapest")  # also: "fastest", "most_reliable", "balanced"
g.get("https://example.com")     # picks via /v1/health/upstreams under the hood

Cost guardrails

g = TierProxy(
    monthly_budget_usd=200.0,    # raises BudgetExceededError before going over
)

Power-user knobs

import httpx
from tierproxy import TierProxy
from tierproxy.retry import RetryPolicy

g = TierProxy(
    api_key="tp_live_...",
    base_url="https://my-self-hosted-gw:8444",
    timeout=10.0,
    retry_policy=RetryPolicy(max_retries=5, retry_on_status=frozenset({500, 502})),
    http_client=httpx.Client(verify=False),  # custom transport
    user_agent_suffix="my-app/2.3",          # attribution
)

Raw modes (Playwright, curl, etc.)

from tierproxy import ProxyURL

p = ProxyURL(api_key="tp_live_...", country="US", mode="username_encoding")
print(p.http_url())  # http://customer-tp_live_...-cc-US:[email protected]:443

Error handling

Every SDK error inherits from tierproxy.TierProxyError and carries a request_id for support escalation:

from tierproxy import TierProxy, RateLimitError
import time

with TierProxy() as g:
    try:
        resp = g.get("https://example.com/page")
    except RateLimitError as e:
        time.sleep(e.retry_after or 5)
        resp = g.get("https://example.com/page")

See Errors reference for the full HTTP-status-to-exception mapping.

AI agent integration

The SDK exposes its response models as JSON Schema and as pre-built tool definitions for Anthropic Claude and OpenAI function-calling:

import anthropic
from tierproxy import TierProxy, schemas

with TierProxy() as gw:
    anthropic.Anthropic().messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        tools=schemas.anthropic_tools(),
        messages=[{"role": "user", "content": "How much quota is left?"}],
    )

See the AI integration guide and the MCP server in examples/mcp_claude_desktop.md.

How tierproxy compares

	tierproxy	Smartproxy SDK	Bright Data SDK	Oxylabs SDK	DataImpulse
Multi-provider routing	✅	❌	❌	❌	❌
Client-side smart selector (cost-aware)	✅	❌	❌	❌	❌
Live usage streaming (SSE)	✅	❌	❌	❌	❌
MCP server (Claude/Cursor/Cline)	✅	❌	❌	❌	❌
OpenTelemetry built-in	✅	❌	❌	❌	❌
Sync + async parity	✅	partial	partial	partial	partial
AI/ML framework examples shipped	8	0	1	0	0
Type-safe (Pydantic v2 + mypy strict)	✅	❌	❌	partial	❌
OpenAPI 3.1 spec	✅	❌	❌	❌	❌
Pip-installable CLI (tierproxy doctor)	✅	❌	❌	❌	❌
Per-request cost attribution (lazy)	✅	❌	❌	❌	❌
JA3/JA4 TLS fingerprint rotation	✅	❌	❌	❌	❌
Rate-limit learning + auto-failover	✅	❌	❌	❌	❌
License	Apache 2.0	proprietary	proprietary	proprietary	proprietary

Features

• Five-second quickstart — import tierproxy; tierproxy.get(url, country="US")
• Layered API — five integration levels from one-liner to power-user knobs
• Smart routing — routing="cheapest" auto-picks healthy upstream per request
• Cost guardrails — monthly_budget_usd= refuses requests that would exceed budget
• Per-request cost attribution — client.cost_for(resp) returns USD; lazy 30s cache, no per-request overhead
• Client-side response caching — cache_ttl=300, cache_max_response_size=262144 LRU with size cap
• Multi-provider auto-failover — auto_failover=True retries with next-best upstream on 429/5xx
• Rate-limit learning — client.rate_limits.get() surfaces gateway-aggregated 429s per target domain
• JA3/JA4 TLS rotation — per-upstream fingerprint randomization (gateway side; see tls-fingerprint guide)
• Cookie persistence — cookies stick to session_id across multi-step crawls
• Streaming responses — client.get(url, stream=True) returns iterator (large files, SSE)
• Live SSE stream — for delta in g.usage.stream() tails month-to-date bytes
• MCP server — tierproxy-mcp exposes proxy as tools to Claude/Cursor/Cline
• 8 framework integrations — LangChain, LlamaIndex, Crawl4AI, Playwright, Firecrawl, Browser-Use, CrewAI
• OpenTelemetry opt-in — pip install tierproxy[otel] for distributed tracing
• Geo + sticky sessions — countries, cities, 1-1440min session pins
• Dual URL syntax — headers (httpx/requests) AND username-encoding (Playwright)
• Type-safe end-to-end — Pydantic v2 models, mypy strict, full IDE autocomplete

See examples/ for LangChain/LlamaIndex/Crawl4AI/Playwright and examples/levels.py for a runnable demo of every level.

Use with your favorite AI/agent framework

Framework	Example	Notes
LangChain	with_langchain.py	RAG document loaders through proxy
LlamaIndex	with_llamaindex.py	SimpleWebPageReader through proxy
Crawl4AI	with_crawl4ai.py	Playwright crawler + tierproxy
Firecrawl (hot)	with_firecrawl.py	Self-hosted Firecrawl + residential IPs
Browser-Use (hot)	with_browser_use.py	LLM-driven autonomous browser
CrewAI (hot)	with_crewai.py	Multi-agent scraper crew + cost-aware routing
Playwright	with_playwright.py	Direct Playwright with tierproxy
MCP (Claude/Cursor/Cline/Windsurf) (unique)	mcp_claude_desktop.md	Native tool integration via tierproxy-mcp

MCP server (Claude Desktop / Cursor / Cline / Windsurf)

pip install tierproxy[mcp]

Then add to your MCP client config:

{
  "mcpServers": {
    "tierproxy": {
      "command": "tierproxy-mcp",
      "env": { "TIERPROXY_API_KEY": "tp_live_..." }
    }
  }
}

Now your AI assistant can call fetch_url(url, country="US"), inspect health and usage, and route through the cheapest healthy upstream — no glue code, no httpx imports, no boilerplate.