Endpoints and model ids

Every request goes to one of two base URLs with a Kano Proxy key, and names its model as provider/model.

Base URLs

Protocol	Base URL	The client appends
OpenAI-compatible	https://<your-domain>/openai/v1	/chat/completions, /responses, /models, /audio/transcriptions
Anthropic Messages	https://<your-domain>/anthropic	/v1/messages, /v1/messages/count_tokens, /v1/models

Some Anthropic SDKs and tools append /messages instead of /v1/messages. For those, set the base to https://<your-domain>/anthropic/v1. Each agent page says which form its tool expects.

The Connect tab on Access shows these URLs for your instance, ready to copy.

Authentication

Send a key from Access in either header. Both headers work on both bases.

Authorization: Bearer <your-api-key>

x-api-key: <your-api-key>

Keys start with sk-kano-proxy-. Your provider credentials stay on the proxy and never reach a client. A key can be revoked on Access at any time, and stops working at once.

Model ids

A model id has two parts, split at the first slash:

<provider>/<upstream model id>

Provider	Example id
claude-code	claude-code/claude-opus-5
codex	codex/gpt-5.6-sol
grok	grok/grok-4.5
antigravity	antigravity/gemini-3-flash
a custom endpoint with slug mygw	mygw/<model the endpoint knows>
a local model with slug my-mac	my-mac/<model the local server reports>

These ids show the shape only. Which models exist for you depends on the accounts you connected. The live list is on Models, or from the API:

curl https://<your-domain>/openai/v1/models -H "Authorization: Bearer <your-api-key>"

One model id works on both bases. You can ask the Anthropic base for a Codex or Gemini model, or the OpenAI base for a Claude model: the proxy converts the request and the response.

An id without a provider prefix is rejected on these bases with 400 invalid_model.

Group endpoints

A model group is a private endpoint with model names you choose. Each name points to one or more provider/model targets, tried in order. Use one when a tool sends fixed model names, or to fall back from one provider to another. Groups explains how to set one up.

Protocol	Group base URL
OpenAI-compatible	https://<your-domain>/g/<group-slug>/openai/v1
Anthropic Messages	https://<your-domain>/g/<group-slug>/anthropic

A group endpoint accepts only that group's model names. GET .../models on it lists exactly those names.

Responses API

/responses is the OpenAI Responses API, the protocol Codex CLI uses. It works with every model id: Codex models pass through unchanged, other providers are converted. Features that need the server to keep state are not available: previous_response_id, conversation, background, and fetching a stored response. Hosted tools such as web search work only on Codex models. See Codex CLI.

Streaming and tools

Streaming (stream: true, or the Anthropic event stream) is passed through as it arrives. Tool calls, images, JSON output modes, stop, temperature, and top_p are forwarded where the target provider supports them.

A streamed response has already sent status 200 when it starts. A failure after that arrives as a final error event in the stream. See Errors.

Reasoning effort

Set it with reasoning_effort on /chat/completions, reasoning.effort on /responses, and output_config.effort or reasoning_effort on the Anthropic base. Accepted values, lowest to highest:

none  low  medium  high  xhigh  max

none turns reasoning off on claude-code, grok and antigravity. codex has no off switch, so the field is left out and the model uses its default. Any other value is rejected with 400 invalid_reasoning. A value above what the provider accepts is lowered to the provider's highest level instead of failing:

Provider	Highest effort
claude-code	max
codex	xhigh
grok	xhigh
antigravity	high
custom endpoint, OpenAI format	sent as is, never lowered
custom endpoint, Anthropic format	not sent from the OpenAI base; on the Anthropic base your thinking settings pass through unchanged

Audio

• Transcription. POST /openai/v1/audio/transcriptions works only with custom endpoints in OpenAI format. Other providers answer 400 unsupported_modality.
• Audio input in chat. An input_audio part in a /chat/completions message reaches antigravity, grok, and OpenAI-format custom endpoints. claude-code, codex, and Anthropic-format custom endpoints reject it with 400 unsupported_modality. The Anthropic base has no audio input.

Which account answered

A non-streamed response carries two headers naming the provider and model that produced it, even when you called a group:

Header	Example
x-kano-upstream-provider	claude-code
x-kano-upstream-model	claude-opus-5

Streamed responses do not carry them. Logs shows the same information for every request.

Not available

• Embeddings.
• Image generation.
• Audio output, including text to speech.