If you're already calling an OpenAI-compatible API, you're most of the way to Token Factory. The wire shape matches — same client libraries, same request bodies, same response envelopes. Two lines change.
This page covers the swap, the model-ID shift, the parameters that don't yet round-trip, and where to find the rest.
#The two-line change
Point your existing OpenAI-compatible client at https://api.tokenfactory.omniva.com/v1 and pass a Token Factory API key.
Mint a Token Factory key from API keys and export it as OMNIVA_API_KEY. Full minting and rotation guidance lives in Authentication.
#Model IDs
OpenAI uses opaque short names — gpt-4, text-embedding-3-small. Token Factory ids are explicit and prefixed, in the form provider/model-name, and the prefix carries meaning. An Omniva/… id (for example Omniva/Kimi-K2.6) is an Omniva-optimized build of an open model — tuned and quantized for low-latency serving. The same base model may also be offered as raw upstream open weights under its author prefix (for example MiniMaxAI/MiniMax-M2.5). Pass the exact catalog id as model.
Pick a model ID from your workspace catalog in Models overview, or list them programmatically with GET /v1/models. The ID you choose goes straight into the model field.
#Parameters dropped today
The chat-completions handler currently reads model, messages, temperature, max_tokens, and stream. The following OpenAI request fields are accepted in the body but ignored — they will not change the response:
top_p,n,seedpresence_penalty,frequency_penaltystoptools,tool_choice,response_formatuser,logprobs,service_tier
If your existing code sets any of these, the call will succeed but those knobs won't take effect. Full list with status is on the Chat completions reference.
#Errors
OpenAI returns a nested envelope ({ "error": { "message", "type", "param", "code" } }). Token Factory returns the same nested shape on the public API surface, but inspect the status and code fields with care — the catalog of codes is Token Factory's, not OpenAI's. The full error catalog and retry guidance is in Errors.
If you're on a corporate network and your first call hits CERTIFICATE_VERIFY_FAILED, see Errors → SSL for the proxy / CA-bundle remediation.
#What about embeddings? Tools? Vision?
POST /v1/embeddings works the same way — swap base URL and pick an embedding model.
Tool-calling round-trip is on the roadmap — see status in the chat-completions guide.
Image input in messages[].content uses the same content-array shape as OpenAI.