ContextQMD
Back to UI-TARS-desktop

SDK Reference

multimodal/websites/tarko/docs/en/guide/get-started/sdk.mdx

0.3.08.4 KB
Original Source

SDK Reference

Comprehensive guide to using the @tarko/agent SDK for building production-ready agents.

Installation

bash
npm install @tarko/agent
# or
pnpm add @tarko/agent

Core Imports

typescript
// Main Agent class
import { Agent } from '@tarko/agent';

// Tool definition
import { Tool, z } from '@tarko/agent';

// Type definitions
import type {
  AgentOptions,
  AgentModel
} from '@tarko/agent';

// Utilities
import { getLogger, LogLevel } from '@tarko/agent';

Agent Class

Constructor

typescript
const agent = new Agent(options: AgentOptions)

AgentOptions Interface

Based on the actual interface from multimodal/tarko/agent-interface/src/agent-options.ts:

typescript
interface AgentOptions {
  // Core Configuration
  id?: string;                    // Unique agent ID (default: "@tarko/agent")
  name?: string;                  // Agent identifier (default: "Anonymous")
  instructions?: string;          // System prompt
  
  // Model Configuration
  model?: AgentModel;             // Model configuration
  maxTokens?: number;            // Token limit per request (default: 1000)
  temperature?: number;          // LLM temperature (default: 0.7)
  top_p?: number;               // Top-p sampling
  thinking?: LLMReasoningOptions; // Reasoning options
  
  // Tools
  tools?: Tool[];                 // Available tools
  tool?: AgentToolFilterOptions;  // Tool filtering options
  toolCallEngine?: ToolCallEngineType; // Tool call engine type (default: 'native')
  
  // Execution Control
  maxIterations?: number;         // Max reasoning loops (default: 1000)
  
  // Context Management
  context?: {
    maxImagesCount?: number;      // Max images in context
  };
  eventStreamOptions?: AgentEventStream.ProcessorOptions;
  enableStreamingToolCallEvents?: boolean; // (default: false)
  initialEvents?: AgentEventStream.Event[]; // Restore events
  
  // Workspace
  workspace?: string;             // Workspace directory
  sandboxUrl?: string;           // Sandbox URL
  
  // Logging
  logLevel?: LogLevel;           // Log level
  metric?: {
    enable?: boolean;            // Enable metric collection (default: false)
  };
}

Core Methods

run() - Execute Agent

Basic Usage:

typescript
// Simple text input - from examples/tool-calls/basic.ts
const response = await agent.run("How's the weather today?");
console.log(response);

With Options:

typescript
// Using AgentRunNonStreamingOptions
const runOptions = {
  input: "How's the weather today?",
};
const response = await agent.run(runOptions);

Streaming:

typescript
// From examples/streaming/tool-calls.ts
const response = await agent.run({
  input: "How's the weather today?",
  stream: true,
});

for await (const chunk of response) {
  console.log(chunk);
}

Tool Definition

Tools are defined using the Tool class and passed to the Agent constructor:

typescript
// From examples/tool-calls/basic.ts
const weatherTool = new Tool({
  id: 'getWeather',
  description: 'Get weather information for a specified location',
  parameters: z.object({
    location: z.string().describe('Location name, such as city name'),
  }),
  function: async (input) => {
    const { location } = input;
    return {
      location,
      temperature: '70°F (21°C)',
      condition: 'Sunny',
      precipitation: '10%',
      humidity: '45%',
      wind: '5 mph',
    };
  },
});

// Tools are passed to Agent constructor
const agent = new Agent({
  tools: [weatherTool],
  // ... other options
});

Tool Definition

Tool Class

Based on the actual Tool class from multimodal/tarko/agent-interface/src/tool.ts:

typescript
const tool = new Tool({
  id: string;                     // Unique tool identifier
  description: string;            // What the tool does
  parameters: ZodSchema | JSONSchema7; // Zod schema or JSON schema
  function: (input: TParams) => Promise<any> | any; // Implementation function
});

Real Examples from Source Code

Location Tool:

typescript
// From examples/tool-calls/basic.ts
const locationTool = new Tool({
  id: 'getCurrentLocation',
  description: "Get user's current location",
  parameters: z.object({}),
  function: async () => {
    return { location: 'Boston' };
  },
});

Weather Tool:

typescript
// From examples/tool-calls/basic.ts
const weatherTool = new Tool({
  id: 'getWeather',
  description: 'Get weather information for a specified location',
  parameters: z.object({
    location: z.string().describe('Location name, such as city name'),
  }),
  function: async (input) => {
    const { location } = input;
    return {
      location,
      temperature: '70°F (21°C)',
      condition: 'Sunny',
      precipitation: '10%',
      humidity: '45%',
      wind: '5 mph',
    };
  },
});

Event Stream

Event Types

Based on multimodal/tarko/agent-interface/src/agent-event-stream.ts, the Agent uses an internal event stream system. Events are emitted during agent execution and can be accessed through streaming responses.

Streaming Events

typescript
// From examples/streaming/tool-calls.ts
const response = await agent.run({
  input: "How's the weather today?",
  stream: true,
});

// Iterate through streaming events
for await (const chunk of response) {
  console.log('Event:', chunk);
  // Events include: user_message, assistant_message_chunk, tool_call, tool_result, etc.
}

Advanced Configuration

Tool Call Engines

typescript
// From actual source code - ToolCallEngineType options
const agent = new Agent({
  toolCallEngine: 'native',              // Use model's native tool calling
  // or
  toolCallEngine: 'prompt_engineering',  // Use prompt-based tool calling
  // or
  toolCallEngine: 'structured_outputs'   // Use structured output parsing
});

Context Management

typescript
const agent = new Agent({
  context: {
    maxImagesCount: 10,           // Max images to keep in context
  },
  maxIterations: 20,             // Allow more reasoning loops
  maxTokens: 4000,              // Limit response length
  temperature: 0.7,             // Control randomness
});

Model Configuration

typescript
// Different model providers
const agent = new Agent({
  model: {
    provider: 'volcengine',
    id: 'doubao-seed-1-6-vision-250815',
    apiKey: process.env.ARK_API_KEY,
  },
  // or
  model: {
    provider: 'openai',
    id: 'gpt-4o',
    apiKey: process.env.OPENAI_API_KEY,
  },
});

Error Handling

typescript
try {
  const response = await agent.run('Complex task');
  console.log(response);
} catch (error) {
  console.error('Agent execution error:', error);
}

Real Examples

Based on the actual examples in multimodal/tarko/agent/examples/:

Basic Tool Calls

typescript
// From examples/tool-calls/basic.ts
import { Agent, Tool, z, LogLevel } from '@tarko/agent';

const locationTool = new Tool({
  id: 'getCurrentLocation',
  description: "Get user's current location",
  parameters: z.object({}),
  function: async () => {
    return { location: 'Boston' };
  },
});

const weatherTool = new Tool({
  id: 'getWeather',
  description: 'Get weather information for a specified location',
  parameters: z.object({
    location: z.string().describe('Location name, such as city name'),
  }),
  function: async (input) => {
    const { location } = input;
    return {
      location,
      temperature: '70°F (21°C)',
      condition: 'Sunny',
      precipitation: '10%',
      humidity: '45%',
      wind: '5 mph',
    };
  },
});

const agent = new Agent({
  model: {
    provider: 'volcengine',
    id: 'doubao-seed-1-6-vision-250815',
    apiKey: process.env.ARK_API_KEY,
  },
  tools: [locationTool, weatherTool],
  logLevel: LogLevel.DEBUG,
});

const response = await agent.run("How's the weather today?");
console.log(response);

Streaming with Tool Calls

typescript
// From examples/streaming/tool-calls.ts
const agent = new Agent({
  model: {
    provider: 'volcengine',
    id: 'doubao-seed-1-6-vision-250815',
    apiKey: process.env.ARK_API_KEY,
  },
  tools: [locationTool, weatherTool],
  toolCallEngine: 'native',
  enableStreamingToolCallEvents: true,
});

const response = await agent.run({
  input: "How's the weather today?",
  stream: true,
});

for await (const chunk of response) {
  console.log(chunk);
}

Next Steps