@mondaydotcomorg/atp-client

@mondaydotcomorg/atp-client

Client library for connecting to Agent Tool Protocol servers and executing code.

Overview

The ATP client enables agents to connect to ATP servers, execute TypeScript code with runtime APIs, handle pauses for LLM/approval/embedding callbacks, and manage client-side tool execution.

Installation

npm install @mondaydotcomorg/atp-client

Architecture

graph TB
    Client[AgentToolProtocolClient] --> Session[ClientSession]
    Client --> API[APIOperations]
    Client --> Exec[ExecutionOperations]
    Client --> Services[ServiceProviders]

    Services --> LLM[LLM Handler]
    Services --> Approval[Approval Handler]
    Services --> Embedding[Embedding Handler]
    Services --> Tools[Client Tools]

    Exec --> Pause[Pause/Resume]
    Session --> Hooks[Pre-Request Hooks]

Quick Start

Basic Client

import { AgentToolProtocolClient } from '@mondaydotcomorg/atp-client';

const client = new AgentToolProtocolClient({
	baseUrl: 'http://localhost:3333',
	headers: {
		Authorization: 'Bearer your-api-key',
	},
});

await client.init();

// Execute code
const result = await client.execute({
	code: `
    const items = ['apple', 'banana', 'cherry'];
    return items.length;
  `,
});

console.log(result.result); // 3

With LLM Support

import { ChatOpenAI } from '@langchain/openai';

const llm = new ChatOpenAI({ modelName: 'gpt-4' });

const client = new AgentToolProtocolClient({
	baseUrl: 'http://localhost:3333',
	headers: { Authorization: 'Bearer key' },
});

// Provide LLM for atp.llm.* calls
client.provideLLM({
	call: async (prompt, options) => {
		const response = await llm.invoke(prompt);
		return response.content;
	},
});

await client.init();

// Now code can use atp.llm.call()
const result = await client.execute({
	code: `
    const joke = await atp.llm.call({
      prompt: 'Tell me a programming joke',
    });
    return joke;
  `,
});

With Embeddings

import { OpenAIEmbeddings } from '@langchain/openai';

const embeddings = new OpenAIEmbeddings();

client.provideEmbedding({
	embed: async (text) => {
		const vector = await embeddings.embedQuery(text);
		// Store and return ID
		return 'embedding-id-123';
	},
	search: async (query, options) => {
		const queryVector = await embeddings.embedQuery(query);
		// Perform similarity search
		return results;
	},
});

// Code can now use atp.embedding.*
const result = await client.execute({
	code: `
    const id = await atp.embedding.embed('Important document');
    const similar = await atp.embedding.search('document', { topK: 5 });
    return similar;
  `,
});

With Approval Handler

client.provideApproval({
	request: async (message, context) => {
		// Prompt user for approval
		const approved = await askUser(message);
		return {
			approved,
			response: { reason: 'User decision' },
			timestamp: Date.now(),
		};
	},
});

// Code can use atp.approval.request()
const result = await client.execute({
	code: `
    const approval = await atp.approval.request(
      'Delete all files?',
      { critical: true }
    );
    
    if (!approval.approved) {
      return { cancelled: true };
    }
    
    return { deleted: true };
  `,
});

Client-Side Tools

Register custom tools that execute on the client:

import { createClientTool } from '@mondaydotcomorg/atp-client';

const fetchTool = createClientTool({
	name: 'fetch',
	description: 'Make HTTP requests from client',
	inputSchema: {
		type: 'object',
		properties: {
			url: { type: 'string' },
			method: { type: 'string', enum: ['GET', 'POST'] },
		},
		required: ['url'],
	},
	handler: async (input) => {
		const response = await fetch(input.url, {
			method: input.method || 'GET',
		});
		return await response.json();
	},
});

const client = new AgentToolProtocolClient({
	baseUrl: 'http://localhost:3333',
	headers: { Authorization: 'Bearer key' },
	serviceProviders: {
		tools: [fetchTool],
	},
});

await client.init(); // Registers tools with server

// Code can now use atp.tool.fetch()
const result = await client.execute({
	code: `
    const data = await atp.tool.fetch({
      url: 'https://api.example.com/data',
      method: 'GET',
    });
    return data;
  `,
});

Pre-Request Hooks

Intercept and modify requests (e.g., token refresh):

const client = new AgentToolProtocolClient({
	baseUrl: 'http://localhost:3333',
	hooks: {
		preRequest: async (context) => {
			// Refresh token if needed
			if (tokenExpired()) {
				const newToken = await refreshToken();
				return {
					headers: {
						...context.currentHeaders,
						Authorization: `Bearer ${newToken}`,
					},
				};
			}
			return {};
		},
	},
});

Core Features

Execute Code

const result = await client.execute({
	code: 'return 1 + 1',
	timeout: 30000,
	pausable: true,
});

// Result structure
interface ExecutionResult {
	status: 'success' | 'error' | 'paused' | 'timeout';
	result?: unknown;
	error?: string;
	pauseReason?: string;
	executionId?: string;
}

Search APIs

const results = await client.search({
	query: 'user authentication',
	limit: 10,
});

// Search by natural language query
const searchResults = await client.searchQuery('how to create a user');

Explore APIs

const schema = await client.explore();

// Returns API structure
interface ExploreResult {
	apis: Array<{
		name: string;
		description: string;
		functions: Array<{
			name: string;
			description: string;
			parameters: unknown;
		}>;
	}>;
}

Type Definitions

const types = await client.getTypeDefinitions();
// Returns TypeScript definitions for atp.* APIs

API Reference

AgentToolProtocolClient

class AgentToolProtocolClient {
	constructor(options: AgentToolProtocolClientOptions);

	// Initialize session
	init(clientInfo?: Record<string, unknown>): Promise<{
		clientId: string;
		token: string;
		expiresAt: number;
		tokenRotateAt: number;
	}>;

	// Execute code
	execute(config: ExecutionConfig): Promise<ExecutionResult>;

	// Resume paused execution
	resume(executionId: string, resumeData: ResumeData): Promise<ExecutionResult>;

	// Provide services
	provideLLM(handler: ClientLLMHandler): void;
	provideApproval(handler: ClientApprovalHandler): void;
	provideEmbedding(handler: ClientEmbeddingHandler): void;

	// API discovery
	search(options: SearchOptions): Promise<SearchResult[]>;
	searchQuery(query: string, options?: SearchOptions): Promise<SearchResult[]>;
	explore(): Promise<ExploreResult>;
	getTypeDefinitions(): Promise<string>;

	// Session info
	getClientId(): string;
}

Options

interface AgentToolProtocolClientOptions {
	baseUrl: string;
	headers?: Record<string, string>;
	serviceProviders?: {
		llm?: ClientLLMHandler;
		approval?: ClientApprovalHandler;
		embedding?: ClientEmbeddingHandler;
		tools?: ClientTool[];
	};
	hooks?: {
		preRequest?: (context: HookContext) => Promise<HookResult>;
	};
}

Flow Diagram

sequenceDiagram
    participant Agent
    participant Client
    participant Server
    participant LLM

    Agent->>Client: execute(code)
    Client->>Server: POST /execute
    Server->>Server: Run code

    alt Code calls atp.llm.call()
        Server->>Client: Pause (LLM callback)
        Client->>Agent: Request LLM
        Agent->>LLM: API call
        LLM-->>Agent: Response
        Agent->>Client: Resume with result
        Client->>Server: POST /resume
        Server->>Server: Continue execution
    end

    Server-->>Client: Execution result
    Client-->>Agent: Return result

Error Handling

import { ExecutionStatus } from '@mondaydotcomorg/atp-client';

try {
	const result = await client.execute({ code: '...' });

	if (result.status === ExecutionStatus.ERROR) {
		console.error('Execution failed:', result.error);
	} else if (result.status === ExecutionStatus.TIMEOUT) {
		console.error('Execution timed out');
	} else if (result.status === ExecutionStatus.PAUSED) {
		console.log('Execution paused:', result.pauseReason);
		// Handle pause (resume later)
	}
} catch (error) {
	console.error('Client error:', error);
}

TypeScript Support

Full TypeScript definitions with strict typing for all APIs.

License

MIT