Generic client example

Many developer tools and agent clients support custom OpenAI-compatible providers. Cosine Inference is designed to fit that pattern.

What you need

To use a generic compatible client with Cosine Inference, you usually need:

• a Cosine bearer token
• a custom provider entry that points the client at https://api.cosine.sh
• the provider wire API set to responses

Base URL

Use:

https://api.cosine.sh

The client should then call the gateway’s /responses endpoint and use /models for model discovery where applicable.

Authentication

Cosine Inference uses a bearer token.

For a typical OpenAI-compatible client, the usual shape is:

• define a custom provider under model_providers
• point env_key at the environment variable that stores your Cosine token
• set base_url to https://api.cosine.sh
• use the Responses wire API

Generic setup pattern

Rather than prescribing one client-specific config format here, use your client’s custom provider documentation and apply these Cosine-specific values:

• provider name: Cosine
• base URL: https://api.cosine.sh
• wire API: responses
• auth: bearer token from your chosen environment variable

Example values

If your client expects the common OpenAI-compatible provider fields, the values to substitute are:

name = "Cosine"
base_url = "https://api.cosine.sh"
wire_api = "responses"
env_key = "COSINE_API_KEY"

Then export your token before launching the client:

export COSINE_API_KEY="your-cosine-token"

Client-specific references

If you want a concrete example for a popular tool, Codex CLI is one compatible option. Use the upstream docs and substitute the Cosine values above:

• Codex config reference
• Codex custom providers

Billing

When you route a compatible client through Cosine Inference, usage is billed through Cosine’s inference gateway rather than through a separate model-provider-specific setup inside the client.

For the higher-level billing model, see Inference Overview.