1. Get a key

Dashboard → API keys → Create. Keys label your tasks (observability). They do not cap concurrency — concurrency is your SKU's max-num-seqs.

2. Point your agent at it

base_url:  https://notoken.cloud/v1
api_key:   tnt-...
model:     qwen3.8-27b

Works with: Cursor, Claude Code (via claude-code-router), Cline, Roo Code, Aider, OpenCode, Windsurf, Zed, Factory Droid, LiteLLM, Portkey, Braintrust, Vercel AI Gateway — anything that speaks OpenAI chat completions.

3. Call it

curl https://notoken.cloud/v1/chat/completions \
  -H "Authorization: Bearer $TNT_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-27b",
    "stream": true,
    "messages": [{"role": "user", "content": "Refactor this module: ..."}]
  }'

Streaming is SSE. Long prefills (big contexts) may pause before the first token — keepalives are injected so proxies (including Cloudflare) don't drop the connection. Hard cap: 10 min per request.

4. Context windows

Native: 262,144 tokens. Extended: up to 1,000,000 (YaRN RoPE scaling, factor 4.0 — Qwen's recommended method). Per-SKU practical limits are on the pricing page. For agentic work, Qwen recommends reserving up to 262K for reasoning + 131K for the final response inside the 1M window.

5. Billing

Prepaid credits (1 credit = $0.01). Your server accrues its SKU rate per second while running. Auto-stop: your configured idle time (default 15 min), or credits = 0 (60s grace so in-flight generations finish). Top up from the dashboard (Stripe) — the server restarts on demand. Per-second billing, no egress fees, no minimums beyond the $10 top-up.

6. What's in every server

Web chat UI (GPU model or CPU model per conversation) · PDF reading · web search while the agent works (DuckDuckGo-weight built-in, BYOK for any search engine) · vision (full mmproj on Fast and Team, CPU vision model on Solo, Lite and Micro).

7. Errors