API reference

Three wire formats, one key, one base URL: https://api.turboprivate.ai.

Authentication

Authorization: Bearer tp_live_…
# or, for clients that send an API key header
x-api-key: tp_live_…

Keys are created and revoked in the app. Revoking takes effect immediately, including for any short-lived token that key has issued.

Short-lived tokens

curl -X POST https://api.turboprivate.ai/v1/token -H "Authorization: Bearer tp_live_…"
# -> { "token": "tps_…", "expires_at": … }  valid for one hour

Useful when a credential has to sit in a config file. The CLI wires this into Claude Code's apiKeyHelper automatically, so a stolen settings file is worthless within the hour.

Endpoints

method	path	format
POST	`/v1/messages`	Anthropic Messages, streaming or not
POST	`/v1/messages/count_tokens`	token estimate
POST	`/v1/responses`	OpenAI Responses
POST	`/v1/chat/completions`	OpenAI Chat Completions
GET	`/v1/models`	what we serve, with prices and context

A first request

curl https://api.turboprivate.ai/v1/chat/completions \
  -H "Authorization: Bearer $TURBOPRIVATE_API_KEY" \
  -H "content-type: application/json" \
  -d '{"model":"ds-flash","messages":[{"role":"user","content":"hi"}]}'

Behaviour worth knowing

Tools work on all three formats, including the full call → result → answer loop.
Unknown model names resolve rather than erroring, so a client that hard-codes a name we do not serve still works.
Streams are kept alive with pings during long prefills, so a client's idle watchdog does not abort a request that is simply thinking.
Errors are passed through with the upstream's own wording and status, because clients decide whether to retry by reading them.
Fields we do not understand are ignored, never rejected — a newer client release cannot break on us.
Limits: requests and concurrency are capped per key. Over the line you get 429 with Retry-After.
No server, no service: if a model has no healthy server we answer 503 and say so, rather than holding your request in a queue.