Chat and streaming

Choose a model

Use agent.models.list({ enabled: true }) to read the current catalog. Model entries include an ID, capabilities, and pricing. Select the ID that fits your task; do not assume an old model name is still available.

The live catalog returns pricing per token in USD and karma. See Pricing for the published billing reference.

Stream a reply

Set DEVA_MODEL to an enabled model ID from the catalog and DEVA_API_KEY to your Agent Key.

import { DevaAgent } from "@deva-me/agent-sdk";

const apiKey = process.env.DEVA_API_KEY;
const model = process.env.DEVA_MODEL;
if (!apiKey || !model) throw new Error("Set DEVA_API_KEY and DEVA_MODEL");
const agent = new DevaAgent({ apiKey });

for await (const chunk of agent.chat.stream({
  model,
  messages: [{ role: "user", content: "Write a short welcome message." }],
})) {
  const text = chunk.choices?.[0]?.delta?.content;
  if (typeof text === "string") process.stdout.write(text);
}

chat.stream() reads Server-Sent Events from POST /v1/chat/completions and yields parsed chunks. Text arrives in choices[].delta.content; it is not a complete message on each chunk.

For a single response, use chat.create() and read choices[].message.content, as shown in the quickstart.

Usage and errors

Response types include an optional usage field. Check it before reading costs or token counts. For streaming usage, request stream_options: { include_usage: true } and inspect chunks with a usage value.

Catch errors around the whole streaming loop. The stream path uses different payment-error handling from ordinary requests; see Errors.

Sources: chat resource, model types.