Chat parameters
Which chat parameters reach the model, which Morse checks, which it refuses, and which it drops.
Chat completions takes OpenAI's parameters. Morse sends some to the model, checks one itself, passes the sampling settings through unless the key is set to refuse them, refuses a few it never sends, and drops the rest without an error, so code written for OpenAI keeps working. The fields on Chat completions are the ones that reach the model.
Sent to the model
| Parameter | Type | What it does |
|---|---|---|
model | string | Required. A chat model's id. Models lists them. |
messages | array | Required. The conversation, at most 4,096 messages. See Messages. |
stream | boolean | Send the reply as server-sent events. Defaults to false. |
stream_options.include_usage | boolean | End a stream with a chunk that carries usage. Without it, a stream carries none. |
max_completion_tokens | integer | The most output tokens the model may write. Below e2ee, Morse lowers a value above the max_output_tokens of the tier that serves the request. Below e2ee, without it the model's default applies; an E2EE request that sets neither it nor max_tokens is refused with e2ee_output_limit_required. |
max_tokens | integer | The same as max_completion_tokens. Below e2ee, max_completion_tokens wins when both are set; at e2ee they must be equal. |
reasoning_effort | string | How hard a reasoning model thinks. See Reasoning effort. |
tools | array | Functions the model may call. See Tool calling. Some TEE and E2EE endpoints can't take them: see Tools on TEE and E2EE. |
tool_choice | string or object | auto, none, required, or one named function. |
parallel_tool_calls | boolean | Whether the model may call several tools in one turn. |
Checked by Morse
provider sets conditions on the model that serves the request. Morse checks them itself and never sends them to the model's provider. See Provider routing.
| Field | What Morse does |
|---|---|
provider.min_privacy | The lowest tier the request may run at: anon, private, tee, or e2ee. A model below it is refused with privacy_floor_unmet, before Morse calls it or holds any credit. |
provider.zdr, provider.data_collection | true, or deny, raises the floor to private and keeps only the endpoints that retain no data. With none left, it's refused with privacy_floor_unmet. |
provider.allow_fallbacks | false keeps only the endpoints at the tier of the first one the request can reach. |
provider.max_price | The most to pay per million tokens, in dollars, as prompt and completion. Morse drops each tier priced above it, and refuses with max_price_unmet when none is left. image, audio, and request are refused with unsupported_parameter. |
provider.quantizations | Keeps only the providers that serve the model at a precision the list names, such as fp8 or bf16, at the tiers the request already reaches. With none left, it's refused with quantization_unmet. |
order, sort, preferred_min_throughput, preferred_max_latency, require_parameters | Accepted and ignored. Morse chooses the provider within a tier. |
only, ignore, enforce_distillable_text | Refused with provider_choice_unsupported. |
| Any other field | Refused with invalid_parameter, which names it. |
Refused when the key turns them off
A new key sends temperature, top_p, stop, seed, and n of 1 to the host as you send them. Turn off Let my settings reach the model host on the key's page in the console, or set forward_settings to false with PATCH /v1/keys/{id}, and a request that sets one of them gets 400 setting_not_forwarded instead, so no setting can tell your requests apart.
Refused on every key
A request gets 400 unsupported_parameter, whatever the key's switch says, when it sets a response_format other than {"type":"text"}, or any of these to a value that changes the reply: logprobs, top_logprobs, logit_bias, frequency_penalty, presence_penalty, top_k, min_p, repetition_penalty, modalities, audio, prediction, verbosity, and web_search_options. A value that changes nothing passes: false, 0, null, an empty object, or modalities of ["text"].
Dropped
Morse drops every other field without an error, and the model's defaults apply. Among them is user.
functions and function_call, OpenAI's older tool fields, are refused with invalid_parameter. Send tools and tool_choice.
An E2EE request is the exception. Morse forwards its body as sent, so it refuses a field outside the E2EE client's short list with e2ee_field_refused, and an output limit above the e2ee tier's max_output_tokens with e2ee_output_limit_refused instead of lowering it, and a request that sets no output limit with e2ee_output_limit_required. See End-to-end encryption.
Messages
Morse keeps the fields of each message that its role uses, and drops name.
| Role | Fields Morse keeps |
|---|---|
system, developer, user | content |
assistant | content, tool_calls |
tool | content, tool_call_id |
content is text, or a list of parts. Media must be inline as a data: URL. An image_url part takes detail set to auto, low, or high. A file part may carry only file_data and filename.
Reasoning effort
reasoning_effort takes none, minimal, low, medium, high, xhigh, or max. Each model lists the efforts it takes as reasoning_efforts in Models, and Morse sends the lowest of them at or above the one you asked for, or the highest it has.
So a model that lists low and high is sent high for medium, and high for max. A model that lists none is sent no effort. See Reasoning.
Other endpoints
Images, search, fetch, and transcription refuse a field they don't take with invalid_parameter, such as size or style on images. Each endpoint's page lists its fields.