API documentation

Connect

Create an account, verify your email, then open the console to create an API key. Share a separate key for each person or application; API keys only authorize inference requests.

Create account · Open console · Community node disclosure

Set CSP_API_BASE_URL to that address, including /v1, and CSP_API_KEY to your key in your server environment. Choose a model from the directory and replace MODEL_ID in examples. The console address may be different from the inference API address.

Send the key in Authorization: Bearer. Keep it on your server; do not put it in client-side JavaScript, URLs or source control. Console login cookies do not authenticate inference requests.

Chat completions

POST /v1/chat/completions returns one completed assistant response as JSON. Streaming is unavailable: omit stream or set it to false. A true value returns HTTP 400 with stream_not_supported.

curl --fail-with-body "${CSP_API_BASE_URL}/chat/completions" \
  -H "Authorization: Bearer ${CSP_API_KEY}" \
  -H 'Content-Type: application/json' \
  --data-binary @- <<'JSON'
{
  "model": "MODEL_ID",
  "messages": [
    {
      "role": "user",
      "content": "Hello."
    }
  ],
  "max_completion_tokens": 256,
  "stream": false
}
JSON

Read the answer from choices[0].message.content. A tool response can instead contain choices[0].message.tool_calls. Check finish_reason; length means the output limit was reached.

Reasoning, unless excluded, is in choices[0].message.reasoning and choices[0].message.reasoning_details.

Responses API, embeddings, image input, strict structured outputs, log probabilities and multiple choices are unavailable.

Request parameters

Use the model directory to check supported parameters. Unknown fields are rejected; input is never silently truncated.

Tools and JSON mode

A tool-capable model returns function names and JSON argument strings. Validate those arguments before executing your application function, then send the assistant tool_calls and the corresponding tool result with its tool_call_id in the next request.

curl --fail-with-body "${CSP_API_BASE_URL}/chat/completions" \
  -H "Authorization: Bearer ${CSP_API_KEY}" \
  -H 'Content-Type: application/json' \
  --data-binary @- <<'JSON'
{
  "model": "MODEL_ID",
  "messages": [
    {
      "role": "user",
      "content": "Hello."
    }
  ],
  "max_completion_tokens": 256,
  "stream": false,
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_time",
        "description": "Read the current time for a time zone.",
        "parameters": {
          "type": "object",
          "properties": {
            "timezone": {
              "type": "string"
            }
          },
          "required": [
            "timezone"
          ]
        }
      }
    }
  ],
  "tool_choice": "auto"
}
JSON

For JSON mode, ask the model for JSON and use response_format. Validate the returned JSON against your application requirements.

curl --fail-with-body "${CSP_API_BASE_URL}/chat/completions" \
  -H "Authorization: Bearer ${CSP_API_KEY}" \
  -H 'Content-Type: application/json' \
  --data-binary @- <<'JSON'
{
  "model": "MODEL_ID",
  "messages": [
    {
      "role": "user",
      "content": "Return a JSON object with a greeting field."
    }
  ],
  "max_completion_tokens": 256,
  "stream": false,
  "response_format": {
    "type": "json_object"
  }
}
JSON

Model list

GET /v1/models requires an API key and returns the gateway's currently usable model list. The website directory publishes catalog definitions and prices; endpoint availability can change independently.

curl --fail-with-body "${CSP_API_BASE_URL}/models" \
  -H "Authorization: Bearer ${CSP_API_KEY}"

Read model IDs from data[].id. Pricing in that API is expressed in dollars per token; the website displays dollars per million tokens.

Usage and billing

The gateway reserves balance using estimated input and the requested output limit before inference. It settles the completed usage and releases the remainder. A balance or API key quota too small for this reservation returns HTTP 402.

Successful responses contain usage.prompt_tokens, usage.completion_tokens and usage.total_tokens. Usage may be estimated when supply reporting is missing or inconsistent. The console's request detail is the billing record.

usage.completion_tokens_details.reasoning_tokens counts reasoning tokens. They are part of completion_tokens and billed as output, even when hidden.

When routing uses a fallback, the response model can differ from your requested model. Billing follows the requested model's price.

Errors and retries

Errors have an error object with type, message, param and code. Save the X-Request-Id response header when requesting support or checking a request in the console.

A client retry is a new inference request and may be charged again. If a connection drops before you receive the result, check the console's usage records before repeating a costly request.