AI - Marina

Declare the service in marina.json, then call marina.ai.generate:

marina.json

{
  "schema": 1,
  "entrypoint": "src/app.ts",
  "runtime": { "ai": "v1" }
}

const result = await marina.ai.generate({
  model: "smart",
  messages: [
    { role: "system", content: "Return a concise factual summary." },
    { role: "user", content: report },
  ],
  maxOutputTokens: 800,
});
return new Response(result.text);

Apps never handle provider credentials or SDKs; Marina routes the call and returns { text }.

Choosing a model

model accepts a stable class or a specific model id:

fast: quick, inexpensive responses.
smart: the stronger default for most work.
A model id from Marina’s library, in vendor/model form, for example anthropic/claude-sonnet-5.

List what your deployment offers, and which account pays, with:

marina ai models

Unknown ids are rejected with the offered list in the error message.

Who pays

By default, generations bill the organization’s Marina-managed key. To bill your own account instead, add an OpenRouter connection under Settings → Connections with your sk-or- API key. While it is active, your workspace’s generations run on your OpenRouter account and its model access.

Limits

Up to 64 messages per request; 32 KiB per message; 128 KiB total input.
maxOutputTokens up to 4,096 (default 1,024 for fast, 2,048 otherwise).
Roles are system, user, and assistant; content is text.
Generation is non-streaming.

Prompts and generated text are never logged or stored by Marina.

Get started

Guides

Platform services

Connections

Reference

AI

Choosing a model

Who pays

Limits

​Choosing a model

​Who pays

​Limits

Choosing a model

Who pays

Limits