WASM Bindings | llms-sdk

WASM bindings for llms-sdk, built with wasm-bindgen. Run OpenAI- and Anthropic-compatible chat completions directly in the browser.

Installation

npm install @cle-does-things/llms-sdk-wasm

The package includes the compiled .wasm module and JS glue. Use it with any bundler that supports WASM (Vite, Webpack, Rollup, etc.) or load it directly in the browser.

Quick start

Import init first to load and instantiate the WASM module, then call chat:

import init, { chat } from "@cle-does-things/llms-sdk-wasm";

await init();

const response = await chat({
  api_type: "openai",
  api_key: "sk-...",
  model: "gpt-5.4-mini",
  messages: [
    textmessage("Hello!")
  ],
  stream: false,
  parallel_tool_calls: false,
});

console.log(response.message.content);

Streaming

Use streamChat with a callback to receive deltas as they arrive:

import init, { streamChat } from "@cle-does-things/llms-sdk-wasm";

await init();

const request = {
  api_type: "openai",
  api_key: "sk-...",
  model: "gpt-5.4-mini",
  messages: [
    textmessage("Count to 5")
  ],
  stream: true,
  parallel_tool_calls: false,
};

await streamChat(request, (err, chunk) => {
  if (err) {
    console.error("Stream error:", err);
    return;
  }

  switch (chunk.type) {
    case "delta":
      console.log("Text:", chunk.delta);
      console.log("Done?", chunk.stop);
      break;
    case "thinkingDelta":
      console.log("Reasoning:", chunk.delta);
      break;
    case "toolDelta":
      console.log("Tool call:", chunk.name, chunk.partial_arguments);
      break;
    case "complete":
      console.log("Final message:", chunk.message);
      console.log("Usage:", chunk.usage);
      break;
  }
});

Multimodal input

Image

import { imagePart } from "@cle-does-things/llms-sdk-wasm";

const message = {
  role: "user",
  content: [
    { type: "text", text: "Describe this image." },
    imagePart("https://example.com/image.png"),     // URL
    // or
    imagePart({ bytes: new Uint8Array([...]) }),    // raw bytes
  ],
};

Document (Anthropic only)

import { documentPart } from "@cle-does-things/llms-sdk-wasm";

const message = {
  role: "user",
  content: [
    { type: "text", text: "Summarize this document." },
    documentPart("https://example.com/doc.pdf"),    // URL
    // or
    documentPart({ bytes: new Uint8Array([...]) }), // raw bytes
  ],
};

Audio (OpenAI only)

import { audioPart } from "@cle-does-things/llms-sdk-wasm";

const message = {
  role: "user",
  content: [
    { type: "text", text: "Transcribe this audio." },
    audioPart({ bytes: new Uint8Array([...]) }),    // raw bytes only
  ],
};

Structured output

Pass an output_format with a JSON Schema to enforce structured responses:

const request = {
  api_type: "openai",
  api_key: "sk-...",
  model: "gpt-5.4-mini",
  messages: [
    { role: "user", content: [{ type: "text", text: "France" }] }
  ],
  stream: false,
  parallel_tool_calls: false,
  output_format: {
    name: "capital",
    description: "Country capital",
    schema: {
      type: "object",
      properties: {
        country: { type: "string" },
        capital: { type: "string" },
      },
      required: ["country", "capital"],
    },
  },
};

Tool use

const request = {
  api_type: "openai",
  api_key: "sk-...",
  model: "gpt-5.4-mini",
  messages: [
    { role: "user", content: [{ type: "text", text: "What's the weather in Paris?" }] }
  ],
  stream: false,
  parallel_tool_calls: false,
  tools: [
    {
      name: "get_weather",
      description: "Return weather for a city.",
      parameters: {
        type: "object",
        properties: {
          city: { type: "string" },
        },
        required: ["city"],
      },
    },
  ],
  tool_choice: "auto",
};

Request schema

Field	Type	Required	Description
`api_type`	`"openai" \| "anthropic"`	✅	Provider type
`base_url`	`string`		Override the default API base URL
`api_key`	`string`	✅	API key
`model`	`string`	✅	Model identifier
`messages`	`Message[]`	✅	Conversation history
`max_output_tokens`	`number`		Maximum tokens to generate
`temperature`	`number`		Sampling temperature
`top_p`	`number`		Nucleus sampling
`reasoning_effort`	`ReasoningEffort`		Control reasoning depth (`none` … `maximum`)
`prompt_cache_ttl`	`string`		Prompt cache TTL hint
`stream`	`boolean`	✅	Enable streaming
`output_format`	`OutputFormat`		Structured output JSON schema
`tools`	`Tool[]`		Available function tools
`tool_choice`	`ToolChoice`		Tool selection mode (`auto`, `none`, `required`)
`parallel_tool_calls`	`boolean`	✅	Allow parallel tool calls