Using the NYU Archway

The Archway is NYU's front door to commercial AI models. NYU holds the vendor accounts; you hold one NYU key beginning sk-nyu-, and that single key reaches every model NYU has enabled. You never see or handle a vendor key.

The gateway speaks the OpenAI API and the Anthropic API, so the SDK you already use keeps working - you change a base URL and a key, and nothing else. It records who called which model and how many tokens, and it enforces a token allowance per person, per vendor.

1. Sign in at the portal.
2. Open Get access and say what you need the key for, and which allowance tier you want. You can have one open request at a time. A demo key needs none of this, and you can use one while you wait.
3. Wait for approval. The decision shows on Get access, and an approved request issues your key in the same moment.
4. Open My keys and press Show me the key. Nobody has seen the secret — not the person who approved you, not the gateway — and you can read it once. Have somewhere to paste it ready: if it is lost you have to request another key. You may hold up to 10 keys at a time.

Two gates govern signing in, and both must pass: the account has to exist, and your address has to be on this deployment's sign-in allowlist. A refused address is deliberately indistinguishable from a wrong password - if you are certain of your password, ask the team whether you are on the list.

Base URL

https://srv1990842.hstgr.cloud/v1 for OpenAI-shaped clients

Base URL

https://srv1990842.hstgr.cloud for the Anthropic SDK, which appends /v1 itself

Your key

Authorization: Bearer sk-nyu-… or x-api-key: sk-nyu-… - both work on every endpoint. If you send both, Authorization wins.

Those endpoints are the whole surface, for every vendor. The dialect you speak and the vendor you reach are independent: an Anthropic-shaped request naming a GPT or Gemini alias is translated for you, and an OpenAI-shaped request naming a Claude alias is too. Use whichever SDK your code already has.

curl https://srv1990842.hstgr.cloud/v1/models \
  -H "Authorization: Bearer $NYU_API_KEY"Copy

Each entry carries the alias to put in model, plus its service, context window and maximum output. Use one of those aliases - a vendor's own model string is not an alias.

GET /v1/providers answers the same question the other way round: one entry per service your key is cleared for, each carrying the aliases it may send, how much of that service's allowance is left, and whether a real vendor is behind it. A service your key cannot use is simply absent. GET /v1/key returns that list plus your key's own scope and expiry, which is the one call to make when a program needs to know everything at once. Neither costs a token, and neither says anything about the account the key belongs to.

export NYU_API_KEY="sk-nyu-…"   # the key you were shown once

curl https://srv1990842.hstgr.cloud/v1/chat/completions \
  -H "Authorization: Bearer $NYU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-fable-5-1",
    "messages": [{"role": "user", "content": "Hello"}],
    "max_tokens": 256
  }'Copy

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["NYU_API_KEY"],
    base_url="https://srv1990842.hstgr.cloud/v1",   # the NYU gateway, not the vendor
)

response = client.chat.completions.create(
    model="claude-fable-5-1",
    messages=[{"role": "user", "content": "Hello"}],
    max_tokens=256,
)
print(response.choices[0].message.content)
print(response.usage)   # the token counts your allowance is measured inCopy

import os
from anthropic import Anthropic

client = Anthropic(
    api_key=os.environ["NYU_API_KEY"],
    base_url="https://srv1990842.hstgr.cloud",   # no /v1 - the SDK appends it
)

message = client.messages.create(
    model="claude-fable-5-1",
    max_tokens=256,
    messages=[{"role": "user", "content": "Hello"}],
)
print(message.content[0].text)Copy

Examples, not the catalogue. GET /v1/models returns exactly what your key may call right now, and is the list to trust.

Your allowance is counted in tokens, per service, per period - spending your Anthropic allowance does not touch your OpenAI one. These headers come back on every response: on success, on denial and on error.

Nothing is zeroed on a schedule. A new period is simply a new counter, so your allowance returns at the moment shown in X-NYU-Quota-Resets. Enforcement here is hard: at the limit, requests are refused.

Burst limits are separate from your allowance and bound abuse rather than cost: 600 requests per minute per key and 1,200 per minute per person. A response carrying X-NYU-Mock: true never reached a vendor - that service has no live credential here, so the content is a deterministic placeholder even though the token accounting is real.

This page is the guide. For the full contract - every endpoint, every request and response field, including the admin and portal JSON APIs - the machine-readable schema is served at /openapi.json, and the repository carries a generated docs/API_REFERENCE.md alongside it.