Use Anthropic clients with Gigatoken

What you need

• •A Gigatoken inference key. Generate one from the Inference console API keys page.
• •A served Gigatoken catalog model, such as qwen/qwen3.5-0.8b. Live availability is exposed by GET /v1/models.
• •The official Anthropic SDK installed in your application.

npm install @anthropic-ai/sdk

Base URL and authentication

export GIGATOKEN_ANTHROPIC_BASE_URL="https://coordinator.gigatoken.baremetallabs.ai"
export GIGATOKEN_INFERENCE_KEY="gt_your_inference_key"
export GIGATOKEN_MODEL="qwen/qwen3.5-0.8b"

Model selection

End-to-end TypeScript sample

This single sample covers text, SDK streaming accumulation, tool use, extended thinking, and token counting.

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  apiKey: process.env.GIGATOKEN_INFERENCE_KEY,
  baseURL:
    process.env.GIGATOKEN_ANTHROPIC_BASE_URL ??
    "https://coordinator.gigatoken.baremetallabs.ai",
});

const model = process.env.GIGATOKEN_MODEL ?? "qwen/qwen3.5-0.8b";

const text = await client.messages.create({
  model,
  max_tokens: 128,
  system: "Reply plainly and briefly.",
  messages: [{ role: "user", content: "Reply with one word: ready" }],
});

const streamed = await client.messages
  .stream({
    model,
    max_tokens: 128,
    system: "Reply plainly and briefly.",
    messages: [{ role: "user", content: "Reply with one word: ready" }],
  })
  .finalMessage();

const weatherTool = {
  name: "get_weather",
  description: "Get weather",
  input_schema: {
    type: "object",
    properties: {
      location: { type: "string" },
    },
    required: ["location"],
  },
} as const;

const toolRequest = await client.messages.create({
  model,
  max_tokens: 128,
  messages: [{ role: "user", content: "What is the weather in Boston?" }],
  tools: [weatherTool],
  tool_choice: { type: "tool", name: "get_weather" },
});

const toolUse = toolRequest.content.find((block) => block.type === "tool_use");
if (!toolUse || toolUse.type !== "tool_use") {
  throw new Error("Expected get_weather tool_use");
}

const toolFollowUp = await client.messages.create({
  model,
  max_tokens: 128,
  messages: [
    { role: "user", content: "What is the weather in Boston?" },
    { role: "assistant", content: toolRequest.content },
    {
      role: "user",
      content: [
        {
          type: "tool_result",
          tool_use_id: toolUse.id,
          content: "72 F and sunny",
        },
      ],
    },
  ],
});

const thinking = await client.messages.create({
  model,
  max_tokens: 2048,
  thinking: { type: "enabled", budget_tokens: 1024 },
  messages: [
    { role: "user", content: "Think briefly, then reply with one word: ready" },
  ],
});

const count = await client.messages.countTokens({
  model,
  messages: [{ role: "user", content: "Reply with one word: ready" }],
});

console.log(text.content);
console.log(streamed.content);
console.log(toolFollowUp.content);
console.log(thinking.content);
console.log(count.input_tokens);

Image input

Supported capabilities

• •Text responses with client.messages.create().
• •Streaming with the SDK accumulator, including client.messages.stream(...).finalMessage().
• •Tool requests, tool_use responses, and follow-up tool_result messages.
• •Extended thinking via thinking: { type: "enabled", budget_tokens: 1024 }.
• •Image input on vision-capable served models.
• •Token counting with client.messages.countTokens().

Known limitations

• •There is no Claude model-name aliasing. Clients must send exact Gigatoken catalog ids.
• •stop_sequences, top_k, and per-block cache_control are accepted, ignored before provider dispatch, and recorded by the adapter's lossy-field tracking.
• •Image sources are limited to base64 and url. Anthropic Files API file_id image sources are not supported.
• •Document and PDF content blocks are currently unsupported.
• •Image input requires a vision-capable model to be currently served.
• •count_tokens rejects images, tools, tool_choice, thinking, document blocks, tool_use, tool_result, and thinking history instead of partially counting unsupported mixed content.

For the general OpenResponses API, continue to use the inference API quickstart.