One endpoint.
Familiar semantics.

OpenAI-compatible by design

Surf exposes a focused subset of the OpenAI API. Existing server-side clients can connect by changing baseURL and using a Surf API key. Requests enter one endpoint while routing stays internal to the network.

Make your first request

1. 01

   Create a key

   Sign in, open API keys, and copy the new key when it is shown. Store it in a server-side secret manager.

2. 02

   Export it locally

   TERMINAL

   export SURF_API_KEY="surf_your_key"

3. 03

   Send a chat completion

   JAVASCRIPT

   import OpenAI from "openai";
   
   const surf = new OpenAI({
     apiKey: process.env.SURF_API_KEY,
     baseURL: "https://route.futurixai.com/v1",
   });
   
   const response = await surf.chat.completions.create({
     model: "kimi-k2.6",
     messages: [{ role: "user", content: "Explain distributed inference." }],
   });
   
   console.log(response.choices[0].message.content);

Bearer keys, kept server-side

Send your key in the Authorization header. Keys are revealed once and stored as hashes. Never expose them in browser code, mobile bundles, or public repositories.

Authorization: Bearer $SURF_API_KEY

Create a completion

curl https://route.futurixai.com/v1/chat/completions \
  -H "Authorization: Bearer $SURF_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k2.6",
    "messages": [{"role":"user","content":"Hello from Surf"}],
    "max_completion_tokens": 4096
  }'

Read output as it arrives

Set stream: true to receive OpenAI-compatible server-sent events. Keep the connection open until the terminal event. Usage is settled after the stream completes or disconnects.

const stream = await surf.chat.completions.create({
  model: "glm-5.2",
  messages: [{ role: "user", content: "Return a short status update." }],
  stream: true,
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}

Let models request your functions

Define tools with JSON Schema. Surf returns tool calls through the standard assistant message shape; your application executes them and sends the result back as a tool message.

{
  "model": "kimi-k2.7-code",
  "messages": [{"role":"user","content":"Check the build status."}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_build_status",
      "description": "Return the current build state",
      "parameters": {
        "type": "object",
        "properties": {"build_id":{"type":"string"}},
        "required": ["build_id"]
      }
    }
  }]
}

Constrain the response shape

Models marked with structured-output support accept response_format. Validate returned JSON in your application even when a schema is supplied.

"response_format": {
  "type": "json_schema",
  "json_schema": {
    "name": "deployment_check",
    "strict": true,
    "schema": {
      "type": "object",
      "properties": {"ready":{"type":"boolean"}},
      "required": ["ready"],
      "additionalProperties": false
    }
  }
}

Discover available aliases

Use the catalog for human-readable capabilities or query the authenticated model endpoint at runtime.

curl https://route.futurixai.com/v1/models \
  -H "Authorization: Bearer $SURF_API_KEY"

Errors use the OpenAI object shape

Inspect the HTTP status and error code. Successful inference responses include x-request-id; log it with your own trace so a request can be located without exposing provider routing identifiers.

{
  "error": {
    "message": "The model 'unknown' does not exist.",
    "type": "invalid_request_error",
    "param": null,
    "code": "model_not_found"
  }
}

Metered against actual tokens

Surf reserves enough credit for the request, then settles against terminal usage. The default output cap is 4,096 tokens and the platform maximum is 32,768. Model context limits are listed in the model catalog; pricing is published per million tokens.