Core Services

Dexto's architecture is built around core services that handle different aspects of agent functionality. Understanding these services helps with debugging, customization, and troubleshooting.

Service Overview

Service	Purpose	Key Responsibilities
DextoAgent	Main orchestrator	Coordinates all services, handles user interactions
MCPManager	Tool coordination	Connects to MCP servers, manages tools and resources
ToolManager	Tool execution	Executes tools, handles confirmations, manages internal tools
SessionManager	Conversation state	Manages chat sessions, conversation history
StorageManager	Data persistence	Handles cache and database storage
PromptManager	System prompts	Manages system prompt assembly and dynamic content
AgentEventBus	Event coordination	Handles inter-service communication

Service Relationships

DextoAgent

Main orchestrator that coordinates all other services.

Key Methods

• start() - Initialize all services
• run(prompt, tools?, sessionId?) - Execute user prompt
• switchLLM(updates) - Change LLM model/provider
• createSession(sessionId?) - Create new chat session
• stop() - Shutdown all services

Usage Example

const agent = new DextoAgent(config);
await agent.start();

// Run a task
const response = await agent.run("List files in current directory");

// Switch models
await agent.switchLLM({ model: "claude-4-sonnet-20250514" });

await agent.stop();

MCPManager

Tool coordination service that connects to Model Context Protocol servers.

Key Methods

• connectServer(name, config) - Connect to MCP server
• disconnectServer(name) - Disconnect server
• getAllTools() - Get all available tools
• executeTool(name, params) - Execute specific tool

Server Types

• stdio - Command-line programs
• http - HTTP REST endpoints
• sse - Server-sent events

Usage Example

// Connect filesystem tools
await agent.mcpManager.connectServer('filesystem', {
  type: 'stdio',
  command: 'npx',
  args: ['-y', '@modelcontextprotocol/server-filesystem', '.']
});

// Get available tools
const tools = await agent.mcpManager.getAllTools();

ToolManager

Tool execution service that handles tool calls and confirmations.

Key Methods

• getToolStats() - Get tool counts (MCP + internal)
• getAllTools() - Get all available tools
• executeTool(call) - Execute tool with confirmation

Tool Confirmation

Controls when users are prompted to approve tool execution:

• auto - Smart approval based on tool risk
• always - Always ask for confirmation
• never - Never ask (auto-approve)

Usage Example

// Get tool statistics
const stats = await agent.toolManager.getToolStats();
console.log(`${stats.total} tools: ${stats.mcp} MCP, ${stats.internal} internal`);

SessionManager

Conversation state management for persistent chat sessions.

Key Methods

• createSession(sessionId?) - Create new session
• loadSession(sessionId) - Load existing session
• listSessions() - List all sessions
• deleteSession(sessionId) - Delete session
• getSessionHistory(sessionId) - Get conversation history

Session Features

• Persistent conversation history
• Session metadata (creation time, last activity)
• Cross-session search capabilities
• Export/import functionality

Usage Example

// Create and switch to new session
const sessionId = await agent.createSession('work-session');

// List all sessions
const sessions = await agent.listSessions();

// Get conversation history
const history = await agent.getSessionHistory('work-session');

StorageManager

Data persistence using two-tier architecture.

Storage Tiers

• Cache - Fast, ephemeral (Redis or in-memory)
• Database - Persistent, reliable (PostgreSQL, SQLite)

Backends

Backend	Use Case	Configuration
in-memory	Development, testing	No config needed
sqlite	Single instance, persistence	path: ./data/dexto.db
postgres	Production, scaling	connectionString: $POSTGRES_URL
redis	Fast caching	url: $REDIS_URL

Usage Pattern

storage:
  cache:
    type: redis                    # Fast access
    url: $REDIS_URL
  database:
    type: postgres                 # Persistent storage
    connectionString: $POSTGRES_CONNECTION_STRING

PromptManager

System prompt assembly from multiple contributors.

Contributor Types

• static - Fixed text content
• dynamic - Generated content (e.g., current date/time)
• file - Content from files (.md, .txt)

Priority System

Lower numbers execute first (0 = highest priority).

Usage Example

systemPrompt:
  contributors:
    - id: primary
      type: static
      priority: 0
      content: "You are a helpful AI assistant..."
    - id: dateTime
      type: dynamic
      priority: 10
      source: dateTime
    - id: context
      type: file
      priority: 5
      files: ["./docs/context.md"]

AgentEventBus

Event coordination for inter-service communication.

Event Types

• thinking - AI is processing
• chunk - Streaming response chunk
• toolCall - Tool execution starting
• toolResult - Tool execution completed
• response - Final response ready

Usage Example

agent.agentEventBus.on('toolCall', (event) => {
  console.log(`Executing tool: ${event.toolName}`);
});

agent.agentEventBus.on('response', (event) => {
  console.log(`Response: ${event.content}`);
});

Service Initialization

Services are initialized automatically when DextoAgent.start() is called:

1. Storage - Cache and database connections
2. Events - Event bus setup
3. Prompts - System prompt assembly
4. MCP - Server connections
5. Tools - Tool discovery and validation
6. Sessions - Session management ready

Debugging Services

Log Levels

# Enable debug logging
DEXTO_LOG_LEVEL=debug dexto

# Service-specific debugging
DEXTO_LOG_LEVEL=silly dexto  # Most verbose

Service Health Checks

// Check MCP connections
const connectedServers = agent.mcpManager.getClients();
const failedConnections = agent.mcpManager.getFailedConnections();

// Check tool availability
const toolStats = await agent.toolManager.getToolStats();

// Check storage status
const storageHealth = agent.storageManager.getStatus();

Common Issues

• MCP connection failures - Check command paths, network access
• Storage errors - Verify database/Redis connections
• Tool execution timeouts - Increase timeout in server config
• Session persistence issues - Check database backend health

Service Configuration

Each service can be configured through the agent config:

# MCP server connections
mcpServers:
  filesystem:
    type: stdio
    command: npx
    timeout: 30000

# Storage backends
storage:
  cache:
    type: redis
  database:
    type: postgres

# Session limits
sessions:
  maxSessions: 100
  sessionTTL: 3600000

# Tool confirmation
toolConfirmation:
  mode: auto
  timeout: 30000

See Configuration Guide for complete config options.