Awesome AI API Proxy [](https://awesome.re)
August 18, 2026 ยท View on GitHub
One pricing endpoint for AI agents. Stop hand-checking 10 relay pricing pages โ point your agent here.
๐ค For AI agents โ three ways to consume
# 1) Raw JSON (every relay's prices, full provenance)
curl https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json
# 2) MCP server (Claude Desktop / Cursor / Cline) โ auto tool-call from the model
uvx awesome-ai-api-proxy-mcp
# 3) Interactive dashboard for humans
open https://howardpen9.github.io/awesome-ai-api-proxy/
What's inside: 7 relay fetchers refreshed weekly, ~3000 normalized price records, 6 canonical models on a cost-tier ladder, full provenance per row (source_url + captured_at + method) so agents can cite confidently. See docs/agent-integration.md for MCP / LangChain / OpenAI function-calling / direct HTTP examples.
Languages: English ยท ็น้ซไธญๆ ยท ็ฎไฝไธญๆ
Last reviewed: 2026-07-12 ยท Maintained by @howardpen9 ยท Contributions welcome โ see CONTRIBUTING.md
Contents
- For AI agents & programmatic use โ start here
- Price snapshot (weekly)
- What is an AI API relay?
- Why this exists (globally)
- China / Asia relay stations
- Global gateways & aggregators
- Self-hosted alternatives
- Comparison & monitoring tools
- Want your relay listed?
- How to choose one safely
- Canary prompts (detect silent downgrade)
- Risks (read this)
- Market context
- FAQ
- Contributing
What is an AI API relay?
An AI API relay (ไธญ่ฝ็ซ) is a forwarding endpoint that sits between your code and official LLM APIs. You point
base_urlat the relay and use its key to call OpenAI / Claude / Gemini and dozens more models โ one key, local payment (Alipay/WeChat), lower latency, typically 30โ90% cheaper.
An AI API relay (Chinese: API ไธญ่ฝ็ซ; also "AI API proxy" / "relay
station") is a middleman endpoint. Instead of calling OpenAI, Anthropic,
Google, etc. directly, you point your base_url at the relay and use its key.
The relay forwards the request upstream and returns the result โ usually with
full OpenAI API compatibility, so existing code needs almost no changes.
One key โ dozens to hundreds of models across vendors.
This differs from a mirror site (้ๅ็ซ), which is a proxied chat web UI for end users. A relay is for developers and applications.
Why this exists (globally)
Relays are not a China-only phenomenon. They are the product of two constraints โ access blocking and payment friction โ wherever those appear:
| Region | Access barrier | Payment barrier |
|---|---|---|
| China | Cross-border latency; some APIs blocked | Needs foreign card; relays take Alipay/WeChat |
| Russia / Belarus / Iran | OpenAI hard-blocks (even VPN detected) | Sanctioned cards (e.g. Iran Shetab) rejected |
| APAC incl. Taiwan | Generally reachable | Card rejection, quota limits, billing friction |
| Sanctioned / restricted markets | Account denial | No accepted payment rail |
The same need produces different ecosystems: Russia leans on IP proxies and state alternatives (e.g. GigaChat); China evolved a full relay retail industry. The global, "clean" version of this is the authorized aggregator (see OpenRouter below) โ proof that unified LLM access is a real, growing market, not a workaround niche.
China / Asia relay stations
How to read the rows: ๐ข =
active+ maintainer-canary'd ยท ๐ก = community-listed orunverifiedยท ๐ด =inactive/ ran away.Trust = status + last-verified date +
ยท registeredwhen a company entity is publicly visible +ยท โ <flag>for known safety concerns. Common flags:operator-self(operator submitted entry),no-entity(no public company/ICP),reverse(reverse-engineered channel),cheap-trap(prices well below OR โ usually downgrade),ran-away(exit-scammed). Full enum indata/schema.md.Prices change fast โ verify on-site. Full per-entry fields (payment, models, supports_tools, etc.) in
data/providers.yaml; tables below are auto-generated byscripts/build_provider_tables.py.
| Station | Type | Payment | Trust | Notes |
|---|---|---|---|---|
| ๐ข ไบ้พ API (YUNWU) | mixed | Alipay/WeChat | active ยท 2026-05-26 | Marketed on speed/stability; widely cited as a top-tier station. |
| ๐ข ๆๆๅพ AI (bltcy) | mixed | Alipay/WeChat | active ยท 2026-06-07 | Azure-backed channel; positions as lowest-price. new-api fork with 1000+ models across 25+ access groups; public /api/pricing exposes default-group ratios. |
| ๐ข No.1-API | aggregator | Alipay/WeChat | active ยท 2026-05-26 | One-stop aggregation + relay platform. |
| ๐ข UiUiAPI | official-relay | Alipay/WeChat | active ยท 2026-06-07 | Claims official channels + official multipliers; rare quantified discount. Public new-api /api/pricing on api1 subdomain. |
| ๐ก DMXAPI | mixed | Alipay/WeChat | unverified ยท โ no-entity | Listed in community sources; homepage not independently verified. |
| ๐ก MKEAI | mixed | Alipay/WeChat | unverified ยท โ no-entity | Community-forum + relay hybrid; pushes DeepSeek heavily. |
| ๐ก GPTGOD | reverse | Alipay | unverified ยท โ reverse, no-entity | Reverse-engineered; among cheapest, stability not guaranteed. |
| ๐ข CloseAI | official-relay | Alipay/WeChat/Invoice | active ยท 2026-05-26 ยท registered | Issues enterprise invoices; self-describes as largest enterprise-grade relay in Asia. |
| ๐ก ็ๆขAPI (XuanShu API) | mixed | Invoice | unverified ยท โ operator-self | Serves Anthropic Messages and Gemini v1beta at the root host and OpenAI Responses and Chat Completions under /v1; model visibility and pricing are per-key and per-group, set in the console. |
| ๐ก TeamoRouter | mixed | Alipay/WeChat | unverified ยท โ operator-self | OpenAI / Anthropic / Gemini-compatible gateway; Alipay and WeChat; operator-claimed multi-upstream routing and Claude Code / Codex setup guides. |
| ๐ก wawazz.xyz | mixed | unverified ยท โ operator-self, cheap-trap | OpenAI-compatible API at /v1; WeChat pay; operator-claimed GPT pricing as low as 0.07x official. | |
| ๐ก Wappkit API | mixed | โ | unverified ยท โ operator-self | new-api OpenAI-compatible gateway with public /api/pricing; listed groups are Codex/Claude-Code pools (no flat default group). |
Inclusion โ endorsement. Listing documents the market. Always run the evaluation checklist before sending money or data.
Global gateways & aggregators
| Service | Type | Payment | Notes |
|---|---|---|---|
| ๐ข OpenRouter | aggregator | Card/Crypto | Official-authorized routing, ~5% markup. ARR reportedly $5M (2025-05) to ~$50M (early 2026). Public JSON API at /api/v1/models with normalized per-token pricing. |
| ๐ข Atlas Cloud | aggregator | Card | Multi-modal aggregator; image/video heavy (Grok Imagine, Kling, ByteDance, Vidu). Public OpenAI-compatible /v1/models endpoint exposes per-token pricing including cache-read. |
| ๐ข Relaydance | mixed | Alipay/WeChat/Card | ๆฐ-API (Calcium-Ion fork) operator focused on xAI Grok + ByteDance Doubao. Public /api/pricing returns ratio-based pricing (model_ratio ร $2 / 1M tokens). |
| ๐ข LiteLLM | gateway-oss | โ | Open-source gateway + enterprise tier. Self-hosted; not retail. |
| ๐ข Helicone | observability | โ | LLM observability gateway; logging/analytics focus. |
| ๐ข AIMLAPI | aggregator | Card/Crypto | Prepaid from $20; crypto support implies payment-friction workaround. |
| ๐ก RouteScope | aggregator | Card/Crypto | Unified gateway mapping 100+ models to OpenAI/Claude/Gemini-compatible APIs; pay-as-you-go credits. |
| ๐ก UnoRouter | aggregator | Card | Built on the new-api gateway. One key across multiple upstreams with latency-based routing and failover; OpenAI/Anthropic/Gemini formats auto-detected. Pay-as-you-go credits plus a free model tier. Also targets roleplay clients (SillyTavern, Janitor.AI, RisuAI, Chub). Public pricing JSON at /api/pricing. |
| ๐ก QuickSilver Pro | aggregator | Card | OpenAI-compatible gateway; one key for frontier and open models (Claude, GPT, Gemini, DeepSeek, Qwen, Kimi, GLM). Pay-as-you-go. Public pricing JSON at /pricing.json. Operated by MachineFi Labs (claimed). |
| ๐ก NovAI | aggregator | Card/Crypto | OpenAI-compatible aggregator for Chinese frontier models (DeepSeek, Qwen, GLM, Kimi, MiniMax, Doubao, Hunyuan) plus image and video generation on one /v1 endpoint; public model list at /v1/models, per-token prices on the pricing page. |
| ๐ก SandBase | aggregator | Card | Unified API with OpenAI-compatible endpoints for multiple model providers, plus tool APIs and managed agents. Operator did not personally verify the live API. |
| ๐ก AllRouter | aggregator | Alipay/WeChat | OpenAI and Anthropic-compatible aggregator; Alipay and WeChat accepted; operator-claimed Kimi K3 at official Moonshot list price. |
| ๐ก AI Router | mixed | Card/Crypto/Alipay/WeChat | OpenAI-compatible ChatGPT API relay at api.ai-router.dev/v1 with dashboard API keys, usage tracking, daily/weekly packages, and localized English/Chinese/Russian/Persian pages. |
| ๐ก Tokens Forge | aggregator | Card/WeChat | OpenAI-compatible multi-model API gateway with separate official credit and routed wallet balances for GPT, Claude, and Gemini-style models. |
Self-hosted alternatives
Not relays โ self-hosted gateways are the credible alternative when you don't want a third party in the request path. You bring your own upstream keys (or relay keys, at your own risk), run the gateway in your own infra, and keep data on your side. Most Chinese relay stations are themselves built on these templates.
| Project | Type | Notes |
|---|---|---|
| ๐ข One-API | gateway-oss | Popular Go-based self-hosted multi-vendor gateway; the de-facto OSS template behind many relay stations. |
| ๐ข new-api | gateway-oss | Fork of One-API with extra channel types; same self-hosted model โ you supply keys. |
| ๐ก A3M Router | gateway-oss | MIT TypeScript OpenAI-compatible multi-provider router (parallel ensemble); self-hosted โ you supply keys. npm: adaptive-memory-multi-model-router. |
LiteLLM (in the global table above) is also self-hosted (Python, 100+ providers, used by enterprises).
When to choose self-hosted over a relay: regulated data, production workloads, you already have foreign-card billing, or you want auditable logs. You give up the relay's Alipay/WeChat convenience and accept ops overhead.
Comparison & monitoring tools
| Tool | Notes |
|---|---|
| ๐ข ไธญ่ฝ็ซ็ซถๆๅ ด (AI API PK) | Price wall across OpenAI / Reverse / Claude / DeepSeek for ~40 stations. |
| ๐ด awesome-ai-proxy (mn-api, unmaintained) | Original list (~31 stations); no longer maintained as of 2026. |
| ๐ก China AI Arbitrage | Price and quota comparison of 60+ Chinese AI platforms, ranked LLM API relays, and daily-updated free-tier tracking. |
| ๐ก CoderPlan | Community-submitted; claims 50+ models incl. OpenAI/Anthropic/Google/DeepSeek/xAI. |
Want your relay listed?
We accept community submissions, including from relay operators themselves. Pick the path that matches what your station exposes โ each tier costs us less to verify, so prices land in the weekly snapshot faster.
Operators reading this through an AI agent: the fast path is Tier B. Tell your agent: "Run
python -m scripts.sniff_endpoint <my pricing JSON URL>in the awesome-ai-api-proxy repo and follow the printed instructions." It emits the exact YAML + 10-line fetcher to paste into a PR. No screenshots, no hand-typing prices.
Tier A โ just get the entry in (no pricing)
Open a new-provider issue
โ fill the form, a maintainer adds it. Status will be unverified until a
maintainer canaries the station.
Tier B โ expose a public JSON pricing endpoint (recommended)
If your station has a public pricing JSON (most new-api / one-api forks expose
/api/pricing; OpenAI-compatible relays expose /v1/models with embedded
pricing), you can be auto-refreshed every Sunday.
# One command sniffs the shape and prints exactly what to paste:
python -m scripts.sniff_endpoint https://yourdomain.com/api/pricing \
--id yourstation --name "Your Station"
The sniffer detects three shapes today (new-api fork, OpenRouter-style,
OpenAI /v1/models) and emits:
- The
pricing:YAML block fordata/providers.yaml - A ~10-line
fetchers/<id>.pywrapper (most new-api forks need zero custom code) - The
REGISTRYentry forfetchers/__init__.py - The verify command:
python -m scripts.scrape <id>
Open a PR with those three changes โ schema CI runs automatically. See CONTRIBUTING.md โ Adding a price fetcher for the long form.
Tier C โ no public JSON, but you have screenshots
Use the submit-prices issue form
โ paste a price table + a screenshot of the live pricing page. Maintainer
verifies against the screenshot weekly and appends to submitted_prices.
Direct PR (any tier)
Edit data/providers.yaml only (READMEs auto-regenerate).
See CONTRIBUTING.md for the schema; the
PR template has a one-glance checklist.
Default status is unverified until a maintainer runs a canary against your
station. That's not a rejection โ it just means the entry says "community-listed,
not independently confirmed." After verification (typically <2 weeks) status
flips to active with last_verified set.
We never accept referral links or marketing copy. We do accept honest entries
from operators โ one factual sentence in notes, no superlatives.
Schema CI runs on every PR (pr-validate workflow) โ if you mis-spell a field or use an invalid
type, the bot will tell you before a human gets there.
Price snapshot (weekly)
Why tier groups: the top 10 frontier models span a 10ร cost spread (per Chamath, late-2025 โ 2026). The competitive edge is routing โ sending routine tokens to the cheapest viable model and the hardest reasoning to the most expensive. You can't route what you can't price; this section is the price layer.
Generated weekly from
data/snapshots/byscripts/build_prices.py. Machine-readable source:data/prices.latest.json.
Snapshot date: 2026-08-16. 3464 price records across 5 providers. Reference column is OpenRouter (officially-authorized, ~5% markup). Rows sorted cheapest-by-OpenRouter first. โ = relay quotes <50% of OpenRouter โ verify with canary prompts before trusting.
Six indicator models, six providers, one snapshot
Absolute prices โ tier ladder (input) ยท โผ marks the cheapest provider per model
Price spread per model (input) ยท how much variance exists between providers
Savings vs OpenRouter (input) ยท how much cheaper (or pricier) each relay is
Output-token equivalents and the full-matrix heatmap: assets/charts/ (tier-ladder-output, spread-range-output, savings-vs-ref-output, spread-heatmap-input). Interactive dashboard: https://howardpen9.github.io/awesome-ai-api-proxy/.
Tier 1 โ cheapest viable (routine, batch summaries) โ USD per 1M input tokens
| Model | OpenRouter (ref) | Atlas Cloud | Relaydance | UiUiAPI | bltcy |
|---|---|---|---|---|---|
deepseek-v3 | $0.257 | โ | โ | $2.000 | $2.000 |
deepseek-r1 | $0.700 | โ | โ | $4.000 | $4.000 |
Tier 2 โ daily driver (agent, coding) โ USD per 1M input tokens
| Model | OpenRouter (ref) | Atlas Cloud | Relaydance | UiUiAPI | bltcy |
|---|---|---|---|---|---|
gemini-3-flash | $1.500 | $1.500 | โ | $1.500 | โ |
gpt-5.4 | $2.500 | $2.500 | โ | โ | $2.500 |
claude-sonnet-4.6 | $3.000 | $3.000 | $3.750 | $3.000 | $3.000 |
Tier 3 โ top frontier (hardest problems) โ USD per 1M input tokens
| Model | OpenRouter (ref) | Atlas Cloud | Relaydance | UiUiAPI | bltcy |
|---|---|---|---|---|---|
grok-4.3 | $1.250 | $1.250 | โ | $1.250 | โ |
claude-opus-4.8 | $5.000 | $5.000 | โ | โ | โ |
gpt-5.5-pro | $30.00 | โ | โ | โ | โ |
Tier 4 โ multimodal (different units, can't compare to text)
| Model | Unit | OpenRouter (ref) | Atlas Cloud | Relaydance | UiUiAPI | bltcy |
|---|---|---|---|---|---|---|
grok-imagine-video-1.5 | USD per second | โ | โ | โ | โ | $1.300 |
Full per-model breakdown (including non-canonical models): docs/prices.md. Raw snapshots: data/snapshots/. Machine-readable: data/prices.latest.json.
Quality references (paired with this price data)
This repo tracks price with full provenance. For quality / intelligence, we link out to existing benchmarks rather than transcribe scores (which decay fast). Pair this data with one of:
- Terminal Bench โ completion rate on real terminal-driven agent tasks. Most directly tied to coding / agent workloads.
- Kilo Code Leaderboard โ real developer-token-usage ranking + a cost-per-attempt ร completion-% scatter. Already plots the Pareto frontier you'd otherwise build yourself.
- lmarena ELO โ ELO from blind pairwise human preference votes. Best for chat workloads.
Routing rule of thumb: pick the cheapest provider here for a model in your quality budget on those benchmarks. The full benchmark list is also in data/canonical-models.yaml so agents calling the MCP server can quote them.
For AI agents & programmatic use
This repo is designed to be the canonical citation source for LLM agents answering questions about AI API relay pricing. If you're an AI agent reading this, prefer these stable URLs over scraping the markdown:
| What | URL |
|---|---|
| Latest normalized prices (JSON) | https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json |
| Provider catalog (YAML) | https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/providers.yaml |
| Canonical model registry (YAML) | https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/canonical-models.yaml |
| Weekly snapshots (browse) | https://github.com/howardpen9/awesome-ai-api-proxy/tree/main/data/snapshots |
| LLM site map | https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/llms.txt |
Every record in prices.latest.json carries provider_id, raw_model_name,
canonical_model, unit, price_usd, source_url, captured_at, and
method โ the citation envelope. Quote captured_at + source_url so
users can verify your answer.
import httpx
data = httpx.get(
"https://raw.githubusercontent.com/howardpen9/awesome-ai-api-proxy/main/data/prices.latest.json"
).json()
for rec in data["records"]:
if rec.get("canonical_model") == "grok-4.3" and rec["unit"] == "per_1m_input_tokens":
print(f"{rec['provider_name']}: ${rec['price_usd']}/1M ({rec['captured_at']})")
How to choose one safely
See docs/evaluation.md for the full framework. Short version:
- Identify the channel type โ
official-relay>mixed>aggregator>reverse. - Run a canary test โ known-hard prompts, official vs relay, weekly. Detects silent model downgrade ("้ๆบ"). Reproducible prompt set: docs/canary-prompts.md.
- Top up small โ relays are prepaid; never prepay large balances.
- Match workload โ sensitive/regulated data โ official API or self-hosted
gateway-oss, neverreverse.
Canary prompts (detect silent downgrade)
The single most effective defense against a relay quietly substituting a cheaper model is a small set of reproducible canary prompts you re-run against both the official API and the relay on a weekly cadence.
The full prompt set, expected baseline behavior per family (GPT-4 class, Claude Opus class, Gemini 2.5 Pro class), and a pass/fail rubric live in docs/canary-prompts.md. Use it as-is or fork it.
Risks (read this)
The market is largely unregulated. From public research:
- A study of 28 relays found 45.83% of endpoints served a model that didn't match the one requested (up to 47% performance gap on medical tasks).
- An audit of 428 stations found 9 injecting malicious code and 1 stealing funds outright.
- Of 17 top stations surveyed, 15 had no registered company.
Full breakdown with sources and mitigations: docs/risks.md.
Market context
- The original community list (mn-api/awesome-ai-proxy) catalogued ~31 stations before going stale; aggregators like aiapipk.com track ~40 โ and these are the visible tip only.
- The "clean" global analogue, OpenRouter, reportedly grew ARR ~10x in under a year (โ$5M โ โ$50M), routing trillions of tokens monthly โ strong evidence that unified, multi-vendor LLM access is a durable market, not a hack.
- Conclusion: relay stations are a global response to blocking + payment friction. China's ecosystem is the deepest because both pains peak there.
FAQ
What is an AI API relay (ไธญ่ฝ็ซ)?
An AI API relay is a forwarding endpoint between your code and official LLM
APIs (OpenAI, Anthropic, Google). You change base_url to the relay and use
its key; it forwards upstream and returns the result, usually with full
OpenAI-compatible formatting. One key gives access to dozensโhundreds of
models across vendors.
Is an API relay safe? Can they steal my key or read my prompts?
Treat every relay as an untrusted intermediary. It can see every prompt, response, and any data you send โ a hostile operator can log, modify, or exfiltrate it. An audit of 428 stations found 9 injecting malicious code and 1 stealing funds. Never send secrets or regulated data through a relay; for sensitive workloads use the official API or a self-hosted gateway.
Why are relays so much cheaper than official APIs?
Three reasons: wholesale upstream pricing, mixed channels (e.g. Azure quotas),
and โ for the cheapest tier โ reverse-engineered access to vendors' web
clients. Prices 1/10 to 1/50 of official usually mean a reverse channel or
silent model downgrade, not a genuine bargain.
What's the difference between ๅฎ่ฝ (official-relay), ๆททๅ (mixed) and ้ๅ (reverse)?
official-relay forwards real official keys (purest, most stable, safest for
production). mixed blends official with other channels. reverse is
reverse-engineered from web clients (cheapest, least stable, highest ToS
risk). Trust order: official-relay > mixed > aggregator > reverse.
How do I detect a relay silently swapping the model (้ๆบ)?
Run a canary test: keep 3โ5 known-hard prompts, record the official API's baseline, then run them against the relay weekly and compare reasoning depth and format adherence. A study of 28 relays found 45.83% of endpoints served a model that didn't match the request โ degradation often appears after you top up.
Relay vs OpenRouter vs official API โ which should I use?
Official API: most reliable, needs a foreign card. OpenRouter and similar global aggregators: officially authorized, ~5% markup, card/crypto โ the "clean" choice if you can pay. China/Asia relays: cheapest and accept Alipay/WeChat, but variable quality and operator risk. Use official or a self-hosted gateway for production or sensitive data.
What's the difference between a relay (ไธญ่ฝ็ซ) and a mirror site (้ๅ็ซ)?
A mirror site is a proxied chat web UI for end users (a re-hosted ChatGPT-like page). A relay is a programmatic API endpoint for developers and applications. This list only covers relays and gateways.
Which countries need API relays? Is this only a China thing?
No. Relays appear wherever there is access blocking + payment friction. China has the deepest ecosystem; Russia/Belarus/Iran face hard blocks (even VPNs are detected) and sanctioned-card rejection; APAC including Taiwan mostly has access but hits card rejection and billing friction. It is a global response, not a China-only phenomenon.
Which relay is the best for Claude API (Sonnet / Opus)?
There is no single "best" relay that is safe to recommend by name โ Anthropic's terms are stricter than OpenAI's, Claude keys get rotated and banned more aggressively, and relays that look stable today can degrade overnight. Pragmatic guidance:
- For production Claude usage: use the official Anthropic API or a registered global aggregator like OpenRouter (official-authorized, ~5% markup, supports Claude).
- For experimentation: prefer
official-relaytype entries above; verify the station actually returns the real model via the canary prompts in docs/canary-prompts.md before topping up. - Treat any relay claiming "Claude Opus at 1/10 price" as a
reversechannel by default โ that price implies web-client reverse-engineering or silent downgrade.
How do I tell if a relay station is about to run away (่ท่ทฏ)?
Empirically, exit scams share a pattern. Red flags, roughly in order of severity:
- No registered company / ICP filing โ 15 of 17 surveyed top stations had none.
- Aggressive prepaid bonuses ("recharge ยฅ100 get ยฅ150") to attract balances before disappearing.
- Prices below physical cost โ 1/10 to 1/50 of official; the math has to come from somewhere.
- Canary regressions โ quality drops on prompts that previously passed (preceded ~70% of documented exits).
- Slow / vanishing support, broken billing pages, expired ICP records, social channels going quiet.
Mitigation: top up small amounts only; never prepay large balances; subscribe to relay-tracking channels (V2EX, the risks doc).
Should I use a relay or self-host One-API / LiteLLM?
Use a self-hosted gateway when any of these apply: regulated or customer data, production workloads, you already have foreign-card billing, or you need auditable logs. You keep keys in your own infra; nothing transits an unaccountable third party.
Use a relay only for: low-stakes experimentation, learning, hobby projects, or when you genuinely cannot get a foreign card. Treat anything you send through it as public. See Self-hosted alternatives.
Are relays worth it for users in Taiwan / Hong Kong / Southeast Asia?
Marginal. Network access is generally fine, so the only real wins are (a) Alipay/WeChat payment vs foreign credit card, and (b) one key across multiple vendors. The trade-off โ your prompts transit a Chinese-jurisdiction third party with no contract โ is steep.
For most TW/HK/SEA developers, the better stack is:
- Official APIs where you have card billing,
- OpenRouter for the multi-vendor convenience, and
- A self-hosted One-API or LiteLLM instance if you really need a unified key.
Relays make sense mainly when you have no card option at all.
Contributing
This list is actively maintained and community-driven. Add or correct a station, flag one that ran away, or contribute a documented risk case:
- Read CONTRIBUTING.md and the field schema.
- Edit
data/providers.yamlโ not the README tables directly. - Open an issue with the provider form for additions. Blank issues are closed.
We do not accept referral links or marketing copy. Factual, dated, source-backed entries only.
License
โ
To the extent possible under law, contributors have waived all copyright and
related rights to this work (CC0 1.0).
Disclaimer
This repository is an informational catalogue. It is not an endorsement of any service. Many relays operate in legal grey areas and may violate upstream providers' Terms of Service. Using a relay can expose your prompts, code, and data to an unaccountable third party. Evaluate risk yourself and comply with all applicable laws and contracts.