OpenAI SDK
Set base_url to the platform address plus /v1 and use a platform key. Works with the Python and Node SDKs and the official CLI.
Docs
Quick start
Point your base_url at TokenPP and call every listed model with a platform key. Bodies match the official protocols, so existing code usually only needs a new base_url and key.
Three steps
From sign-up to your first response: create a key, add credit, change the base_url.
Create a platform key
Sign in and create a key on the console API Keys page. Keys look like sk-tpp-…, are shown in full only once, and are listed by prefix afterwards. Store it somewhere safe right away.
Add credit
Billing is prepaid. Top-ups and redemption codes credit your balance immediately. With a zero balance, requests return 402.
Point base_url at the platform
OpenAI-compatible clients use the platform address plus /v1. Anthropic-compatible clients (Claude Code, Anthropic SDK) use the platform address alone and the SDK appends /v1/messages. The quick-start card on the console overview shows the actual address of your deployment and copies it in one click.
curl https://<platform-host>/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer sk-tpp-..." \
-d '{
"model": "deepseek-v4.1-flash",
"messages": [{ "role": "user", "content": "ping" }]
}'Common clients
Any client that lets you override base_url can be pointed here. These are the usual setups.
Cursor
In Settings → Models, enable OpenAI API Key, paste a platform key, and set Override OpenAI Base URL to the platform address plus /v1.
Claude Code
Set ANTHROPIC_BASE_URL to the platform address without /v1 and ANTHROPIC_AUTH_TOKEN to a platform key.
Your own app
Any HTTP client that can change base_url and the Authorization header works. Keep the body in the official OpenAI or Anthropic shape.
Endpoints
The platform exposes both OpenAI and Anthropic compatible entry points and authenticates them the same way: Authorization: Bearer <platform key>. Anthropic clients may also use the x-api-key header.
| Endpoint | Protocol | Notes |
|---|---|---|
| POST/v1/chat/completions | OpenAI | Chat completions with SSE streaming |
| POST/v1/responses | OpenAI | Responses API with SSE streaming |
| POST/v1/messages | Anthropic | Messages API with SSE streaming |
| GET/v1/models | OpenAI | Lists sellable models with unit prices; models without a price are omitted |
Streaming and timeouts
All three chat endpoints support SSE. Turn it on for long answers.
- With stream: true the response is text/event-stream, frames arrive as data: {...}, and the stream ends with data: [DONE].
- Frames are returned as they arrive. The whole response is not buffered; if the connection drops, it closes.
- If no new data arrives for too long (90 seconds by default), the connection is closed as a timeout. Clients should handle an early close.
Billing and usage
You pay for the tokens actually used, with no subscription tiers.
- Prices are quoted per 1M tokens in USD, with input, output and cache read/write priced separately. See the model pricing page.
- Before a request is sent, the platform estimates a temporary hold from max_tokens (or max_output_tokens) and the output price, settles against the real usage afterwards, and releases the unused part of the hold.
- Streaming chat requests automatically get stream_options.include_usage so the final frame carries usage, and that usage is what gets billed.
- Every call is written to the usage log with model, token breakdown, charged amount, latency and status, all visible on the console usage page.
- Once the balance runs out, requests return 402.
Error codes
Gateway errors follow the OpenAI error shape: error.message, error.type, error.code, error.param.
| HTTP | code | Meaning |
|---|---|---|
| 401 | invalid_api_key | Key missing, invalid or deleted |
| 401 | key_disabled | This key is disabled |
| 401 | key_expired | This key has expired |
| 403 | account_disabled | The account is disabled |
| 402 | insufficient_balance | Balance too low to place the hold |
| 402 | insufficient_quota | This key has used up its quota |
| 400 | invalid_body | Body is not valid JSON, or model is missing |
| 404 | model_not_found | Model does not exist or is not listed |
| 429 | rate_limited | RPM, TPM or concurrency limit for this key was hit |
| 503 | upstream_unavailable | The service is temporarily unavailable. Try again later. |
FAQ
What exactly goes in base_url?
OpenAI-compatible clients use the platform address plus /v1; Anthropic-compatible clients use the platform address alone. The console overview copies the real address of your deployment for you.
Why do I get model not found?
The model is not listed, or has no price configured. Check the model pricing page for the slug.
What should I check on a 401?
Make sure the header is Authorization: Bearer <platform key> and that you are using a platform key starting with sk-tpp-. A disabled or expired key also returns 401.
My stream stopped halfway. What happened?
Usually the connection dropped, or no data arrived long enough to hit the idle timeout. Handle an early close and retry if needed. Usage already generated is still billed and logged.
How do I see what each call cost?
The console usage page lists token breakdown and charged amount per request, grouped by model and key.