Skip to content

atlas-cloud.md

Latest commit

 

History

History
237 lines (193 loc) · 18.9 KB

atlas-cloud.md

File metadata and controls

237 lines (193 loc) · 18.9 KB
Raw
title Atlas Cloud Provider Guide
description Atlas Cloud on NeuroLink — OpenAI-compatible Tier-2 catalog provider, default model `deepseek-ai/deepseek-v3.2`
keywords atlas cloud, atlascloud, openai-compatible, tier 2, provider setup, deepseek, glm, qwen, kimi

Atlas Cloud Provider Guide

Atlas Cloud is a Tier-2 catalog provider: OpenAI-wire-compatible with no behavioural quirks, so its entire integration is one JSON file (src/lib/providers/catalog/atlas-cloud.json) rather than hand-written code. That file is the source of truth for everything on this page.

Verification status: this entry is docs- and roster-verified, not yet live-verified. Every field below comes from Atlas Cloud's own public documentation and an unauthenticated GET /v1/models call — no account was created and no API key was used to build it. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

Key Facts

  • Provider id: atlas-cloud
  • Protocol: OpenAI-compatible (POST /v1/chat/completions, Bearer auth) — Atlas Cloud's LLM API Protocols page (docs/llm-protocols) also lists Anthropic Messages, Responses and Gemini formats on the same base URL; NeuroLink uses only the chat-completions one
  • Base URL: https://api.atlascloud.ai/v1 (the /v1 suffix is required — docs/models/llm)
  • Default model: deepseek-ai/deepseek-v3.2 — Atlas Cloud is an aggregator with no single house model; this is the model its own chat-completions, Anthropic-Messages and Responses examples all use
  • Models in catalog: 10 (curated from a 120-model live roster)
  • Streaming: supported — the LLM / Chat page states the API "supports both streaming and non-streaming modes"
  • Tool calling: model-dependent — the docs say tools, tool_choice and parallel_tool_calls are passed through on models that advertise tools
  • Structured output: supported — response_format accepts json_object and json_schema on models that advertise json_mode or structured_outputs; every model in this catalog advertises one of them
  • Structured output + tools together: not declared (false) — no combined probe was run without credentials
  • Tool calls under streaming: not declared (false) — no page describes streamed tool-call deltas
  • Thinking: declared (true) — the default model carries the roster's reasoning feature and the protocols page documents normalised reasoning flags (enable_thinking, thinking.type, reasoning_effort: "none"); not exercised live
  • Embeddings: false — the protocols page lists /v1/embeddings under "Not available"
  • Billing: no-free-tier is the conservative schema value, not a verified fact — see Billing below
  • Key format: starts with apikey- (docs/llm-protocols)

Quick Start

1. Get an API key

  1. Visit: https://www.atlascloud.ai/login and sign up or sign in (Google, GitHub, or email and password — one page handles both, see https://www.atlascloud.ai/docs/account/sign-in)
  2. Open https://www.atlascloud.ai/console/api-keys, choose Create API Key and copy it immediately — it is shown only once (https://www.atlascloud.ai/docs/api-keys), and Atlas Cloud API keys begin with apikey- (https://www.atlascloud.ai/docs/llm-protocols)
  3. Billing is pay-as-you-go from a prepaid balance: the quick start's Step 3 is Add Funds with a $25 minimum top-up (https://www.atlascloud.ai/docs/get-started), and requests are rejected with 402 while the balance is insufficient (https://www.atlascloud.ai/docs/errors). The homepage FAQ also says 'No credit card required to start', but no free-tier terms are documented, so check your balance at https://www.atlascloud.ai/console/billing before relying on it
  4. Set ATLAS_CLOUD_API_KEY in your .env file (ATLASCLOUD_API_KEY, the name used in Atlas Cloud's own docs, is also read)

2. Configure

export ATLAS_CLOUD_API_KEY=apikey-your-key
export ATLAS_CLOUD_MODEL=deepseek-ai/deepseek-v3.2   # optional — overrides the default model
export ATLAS_CLOUD_BASE_URL=https://api.atlascloud.ai/v1   # optional — proxy or self-hosted gateway

3. Use it

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "atlas-cloud",
  model: "deepseek-ai/deepseek-v3.2",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider atlas-cloud

Per-request credentials work as they do for every provider:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "atlas-cloud",
  credentials: { atlasCloud: { apiKey: process.env.ATLAS_CLOUD_API_KEY } },
});

Models

Model Context Vision $/M in · out Notes
deepseek-ai/deepseek-v3.2 ⭐ 164K no $0.26 / $0.38 (cached $0.13) Default here because the chat-completions, Anthropic Messages and Responses examples on the LLM API Protocols page use it (the vendor does not recommend it); roster tags tools, structured_outputs, reasoning; model page says 128K context, see below.
deepseek-ai/deepseek-v4-flash 1M no $0.14 / $0.28 (cached $0.028) First fallback; 393K output ceiling on the roster.
zai-org/glm-5.3 1M no $1.40 / $4.40 (cached $0.26) Fallback from a second family (GLM is a featured family on the LLM / Chat page); 131K output ceiling.
Qwen/Qwen3-235B-A22B-Instruct-2507 131K no $0.20 / $0.88 (cached $0.20) Fallback from a third family (Qwen is featured on the LLM / Chat page).
deepseek-ai/deepseek-v4-pro 1M no $1.68 / $3.38 (cached $0.13) Roster output ceiling 393K.
zai-org/glm-5.1 203K no $1.26 / $3.96 (cached $0.234) Roster output ceiling equals context (203K), so none recorded separately.
minimaxai/minimax-m2.7 197K no $0.30 / $1.20 (cached $0.06) MiniMax is a featured family on the LLM / Chat page.
moonshotai/kimi-k2.6 262K yes $0.95 / $4.00 (cached $0.16) Roster input modalities text, image, video; models.visionModel.
qwen/qwen3.5-397b-a17b 262K yes $0.55 / $3.50 (cached $0.55) Roster input modalities text, image, video; 65K output ceiling.
qwen/qwen3.5-27b 262K yes $0.27 / $2.16 (cached $0.27) Roster input modalities text, image, video; 65K output ceiling.

Context, pricing, output ceilings and modalities are taken from an unauthenticated GET https://api.atlascloud.ai/v1/models call made 2026-09-29 (120 models total; 10 curated into this catalog, all of them text-output models whose roster entry advertises tools plus json_mode or structured_outputs). The roster reports prices per token; the table multiplies by one million. Where a model's max_output_length equals its context window, no separate output ceiling is recorded in the catalog. models.defaultContextWindow and models.defaultMaxOutputTokens (both 163,840) are the default model's own roster figures.

Default model's context window — the vendor's pages disagree. The model page for deepseek-ai/deepseek-v3.2 (models/deepseek-ai/deepseek-v3.2) says "128K token context length support" and lists "Context Length: 128K tokens" in its spec table. The roster's context_length and the same page's machine-readable reference (llms.txt: "Context length: 163,840 tokens", "Max output tokens: 163,840") both say 163,840. The catalog records 163,840 because the two API-facing sources agree on it; the vendor does not say which figure governs. No request longer than 128K tokens was sent, so treat the range between 128K and 163,840 as vendor-declared and untested; if you need certainty, keep prompts under 128K tokens.

Fallback order when the default is unavailable: deepseek-ai/deepseek-v4-flash → zai-org/glm-5.3 → Qwen/Qwen3-235B-A22B-Instruct-2507.

Vision tags (vision: true) come from the image entry in each model's roster input_modalities, not from an image-input probe — no key was used to send a real image request. Image parts use the OpenAI image_url content-part shape described on the protocols page.

About is_ready. In the 2026-09-29 roster call the is_ready field was false on 73 of the 120 entries (by id prefix: anthropic/ 29, openai/ 18, google/ 11, qwen/ 8, xai/ 4, moonshotai/ 2, zai-org/ 1) and absent on the other 47. The models in this catalog are entries where it is absent. The roster also comes back in a custom {code, msg, data} envelope rather than OpenAI's {object: "list"} wrapper, but data is still an array of {id, ...} entries.

Billing

The vendor's pages do not agree, and the catalog does not paper over that:

  • The billing docs (docs/billing, top-up, balance & credits) and the quick start describe pay-as-you-go from a prepaid balance with a $25 minimum top-up; a request made with an insufficient balance is rejected with HTTP 402 (docs/errors). Bonus credits exist (promotions, redeem codes, referrals) but are not described as a free tier.
  • The homepage FAQ says "No credit card required to start" in its answer to "How do I get started"; the phrase "free credits" appears only in the homepage blurbs for the GPT Image 2 and Nano Banana 2 models, with no terms. The pricing page is headed "Pay Per Use, No Subscriptions".

No page states what a brand-new, unfunded account can actually call, so billingPolicy is set to no-free-tier because the schema has no "unknown" value and that is the conservative choice for a setup wizard — not because it was confirmed. Check the balance in the console before relying on it.

Verification status

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for Atlas Cloud — docs- and roster-verified, not live-verified:

Probe Result
Roster unauthenticated GET /v1/models, HTTP 200, 120 models, 2026-09-29 — no key used
Billing public docs, quick start and homepage read 2026-09-29; prepaid balance and 402 on insufficient balance are documented, the homepage FAQ's "No credit card required to start" is not reconciled with them — see Billing
Auth-failure shape documented on docs/errors ({"code":401,"msg":"unauthorized","request_id":"…","data":null}; a wrong path also returns 401; 402 is insufficient balance; 404 covers models not available to the account) and not probed live — errorRules is built from that page alone
Tools / structured output documented on docs/llm-protocols for models that advertise them (the page does not say what happens on other models); neither was exercised live, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweep not run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=atlas-cloud with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting

Symptom Cause Fix
HTTP 401 on every request Key missing, unknown or expired — or a mistyped endpoint path (the gateway authenticates first) Check ATLAS_CLOUD_API_KEY, and confirm the base URL ends in /v1
HTTP 402 Insufficient balance Add funds in the console; service resumes after top-up
Model not found (404) Resource not found — the docs say a 404 for a model also covers models that are not available to your account (docs/errors) Check the model id against the catalog; the unauthenticated GET https://api.atlascloud.ai/v1/models roster lists ids
HTTP 429 with no Retry-After Per-account, per-model rate limit; LLM endpoints send no X-RateLimit-* or Retry-After headers Back off client-side with jitter (docs/errors)
200 on a stream, but the reply is cut short Errors after a stream opens arrive as events inside the stream Handle mid-stream termination; ignore SSE comment lines starting with : (keep-alives)
Tool calls or response_format misbehave The docs cover them for models that advertise tools / json_mode / structured_outputs; behaviour on other models is undocumented Use a model from the table above (all advertise tools and json_mode or structured_outputs)
Structured output ignored with tools attached structuredOutputWithTools is false on this entry — untested combination NeuroLink omits response_format automatically whenever tools are present, before sending
Reply contains a system prompt you never sent The gateway inserts "You are a helpful assistant." when no system prompt is sent Send your own system message
Embeddings call fails Atlas Cloud has no /v1/embeddings endpoint Use a different provider for embeddings

See also