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

WASM bindings for llms-sdk, to call OpenAI- and Anthropic-compatible LLMs from the browser.

Installation

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

Or use directly in a browser via a bundler that supports WASM (e.g. Vite, Webpack, Rollup).

Quick Start

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: [
    { role: "user", content: [{ type: "text", text: "Hello!" }] }
  ],
  stream: false,
  parallel_tool_calls: false,
});

console.log(response.message.content);

API

Initialization

import init from "@cle-does-things/llms-sdk-wasm";
await init(); // loads and instantiates the WASM module

chat(request, retry_policy?)

Sends a single non-streaming chat request and returns the full response.

Parameters:

• request: LLMRequest — the chat request (see LLMRequest)
• support_developer: boolean — Whether or not the OpenAI-like clients support developer as a role instead of system.

Returns: Promise<LLMResponse>

const response = await chat({
  api_type: "openai",
  base_url: "https://api.openai.com/v1",  // optional
  api_key: "sk-...",
  model: "gpt-5.4-mini",
  messages: [
    { role: "system", content: [{ type: "text", text: "You are a helpful assistant." }] },
    { role: "user", content: [{ type: "text", text: "Explain quantum computing." }] }
  ],
  max_output_tokens: 256,
  temperature: 0.7,
  top_p: 1.0,
  stream: false,
  parallel_tool_calls: false,
});

// response.id          — provider-generated response ID
// response.created_at  — unix timestamp
// response.message     — { role, content: MessagePart[] }
// response.usage       — { input_tokens, output_tokens, cache_read_tokens, cache_write_tokens }

streamChat(request, callback, retry_policy?)

Streams a chat response, invoking the callback for each chunk.

Parameters:

• request: LLMRequest — the chat request
• support_developer: boolean — Whether or not the OpenAI-like clients support developer as a role instead of system.
• callback: (error: any, chunk: LLMStreamingResponse) => void — called on every stream event

Returns: Promise<void>

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;
  }
});

Content Helpers

Build MessagePart objects for multimodal inputs:

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

// Image from URL or Uint8Array
const img = imagePart("https://example.com/image.png");
const imgFromBytes = imagePart({ bytes: new Uint8Array([...]) });

// Document (PDF, etc.) from URL or Uint8Array
const doc = documentPart("https://example.com/doc.pdf");

// Audio from Uint8Array
const audio = audioPart({ bytes: new Uint8Array([...]) });

Types

LLMRequest

| 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 |

Message

interface Message {
  role: "user" | "assistant" | "tool" | "system";
  content: MessagePart[];
}

MessagePart

A discriminated union of content types:

| Tag | Fields | |-----|--------| | { type: "text" } | text: string | | { type: "image" } | image_data: string, is_base64: boolean, mime_type?: string | | { type: "document" } | document_data: string, is_base64: boolean, mime_type?: string | | { type: "audio" } | audio_data: string, mime_type: string | | { type: "toolCall" } | id: string, name: string, arguments: string | | { type: "toolResult" } | tool_call_id: string, result: string | | { type: "thinking" } | thinking: string, signature?: string |

LLMResponse

interface LLMResponse {
  id: string;
  created_at?: number;
  message: Message;
  usage: LLMUsage;
}

LLMUsage

interface LLMUsage {
  input_tokens: number;
  output_tokens: number;
  cache_read_tokens?: number;
  cache_write_tokens?: number;
  other_tokens?: Map<string, number>;
}

Streaming Types

type LLMStreamingResponse =
  | { type: "delta"; response_id: string; created_at?: number; delta?: string; stop: boolean }
  | { type: "thinkingDelta"; response_id: string; created_at?: number; delta?: string }
  | { type: "toolDelta"; response_id: string; tool_call_id: string; name: string; partial_arguments: string }
  | { type: "complete"; id: string; created_at?: number; message: Message; deltas: LLMStreamingDelta[]; thinking_deltas?: LLMThinkingDelta[]; usage?: LLMUsage; tool_calls?: ToolCallPart[] };

Demo

A complete browser demo is included in demo.html. Open it in a browser (via a local server) after building the WASM package:

cd packages/wasm
wasm-pack build --target web
# serve the directory, e.g.
npx serve .

Then open http://localhost:3000/demo.html.

Building from Source

Requires Rust and wasm-pack:

npm run build

The compiled package is output to pkg/.

License

MIT