AxCrew — build a crew of AI agents with shared state (powered by AxLLM)

AxCrew lets you define a team of AI agents in config and run them together with shared state, tools, streaming, MCP, and built‑in metrics/cost tracking. Bring your own functions or use the provided registry.

Why AxCrew

• Config-first crews: Declare agents once; instantiate on demand.
• Shared state: Simple key/value state all agents can read/write.
• Sub‑agents and tools: Compose agents and functions cleanly.
• Streaming: Real‑time token streaming for responsive UX.
• MCP: Connect agents to MCP servers (STDIO, HTTP SSE, Streamable HTTP).
• Metrics & costs: Per‑agent and crew snapshots, with estimated USD.

Install

Install this package:

npm install @amitdeshmukh/ax-crew

AxLLM is a peer dependency, so you will need to install it separately.

npm install @ax-llm/ax @ax-llm/ax-tools

Requirements: Node.js >= 21.

Environment

Set provider keys in your environment. Example .env:

GEMINI_API_KEY=...
ANTHROPIC_API_KEY=...
OPENAI_API_KEY=...

In each agent config, set providerKeyName to the env var name.

Quickstart

This package includes TypeScript declarations and provides type safety.

import { AxCrew, AxCrewFunctions, FunctionRegistryType, StateInstance } from '@amitdeshmukh/ax-crew';
import type { AxFunction } from '@ax-llm/ax';

// Type-safe configuration
const config = {
  crew: [{
    name: "Planner",
    description: "Creates a plan to complete a task",
    signature: "task:string \"a task to be completed\" -> plan:string \"a plan to execute the task\"",
    provider: "google-gemini",
    providerKeyName: "GEMINI_API_KEY",
    ai: {
      model: "gemini-1.5-pro",
      temperature: 0
    }
  }]
};

// Create custom functions with type safety
class MyCustomFunction {
  constructor(private state: Record<string, any>) {}
  
  toFunction(): AxFunction {
    return {
      name: 'MyCustomFunction',
      description: 'Does something useful',
      parameters: {
        type: 'object',
        properties: {
          inputParam: { type: 'string', description: "input to the function" }
        }
      },
      func: async ({ inputParam }) => {
        // Implementation
        return inputParam;
      }
    };
  }
}

// Type-safe function registry
const myFunctions: FunctionRegistryType = {
  MyCustomFunction
};

// Create crew with type checking
const crew = new AxCrew(config, myFunctions);

// Set and get state
crew.state.set('key', 'value');
const value: string = crew.state.get('key');

// Add agents to the crew
await crew.addAgentsToCrew(['Planner']);
const planner = crew.agents?.get('Planner');

if (planner) {
  const response = await planner.forward({ task: "Plan something" });
  
  // Sub-agent usage - when used by another agent (AI is ignored and agent's own config is used)
  const subAgentResponse = await planner.forward(ai, { task: "Plan something" });
  
  // Metrics (per-agent and per-crew)
  const agentMetrics = (planner as any).getMetrics?.();
  const crewMetrics = crew.getCrewMetrics();
  console.log('Agent metrics:', agentMetrics);
  console.log('Crew metrics:', crewMetrics);
}

Key TypeScript features:

• Full type definitions for all classes, methods, and properties
• Type-safe configuration objects
• Proper typing for function registries and custom functions
• Type checking for state management
• Comprehensive type safety for agent operations and responses
• Usage cost tracking with proper types

Concepts at a glance

• Agent: An LLM program with a signature, provider, ai model config, optional examples, and optional mcpServers.
• Sub‑agents: List other agent names in agents to compose behaviors.
• Functions (tools): Register callable functions via a registry and reference by name in agent functions.
• State: crew.state.set/get/getAll() shared across all agents.
• Persona: Use definition (preferred) or prompt to set the system program. If both are present, definition wins.
• Streaming: Use streamingForward() for token streams.
• Metrics: Per‑agent getMetrics() + crew‑level getCrewMetrics() snapshots.

Usage

Initializing a Crew

A Crew is a team of agents that work together to achieve a common goal. You can configure your crew in two ways:

1. Using a JSON configuration file that defines the agents in the crew, along with their individual configurations.
2. Directly passing a JSON object with the crew configuration.

Using a Configuration File

See agentConfig.json for an example configuration file.

// Import the AxCrew class
import { AxCrew } from '@amitdeshmukh/ax-crew';

// Create a new instance of AxCrew using a config file
const configFilePath = './agentConfig.json';
const crew = new AxCrew(configFilePath);

Using a Direct Configuration Object

You can also pass the configuration directly as a JSON object:

// Import the AxCrew class
import { AxCrew } from '@amitdeshmukh/ax-crew';

// Create the configuration object
const config = {
  crew: [
    {
      name: "Planner",
      description: "Creates a plan to complete a task",
      signature: "task:string \"a task to be completed\" -> plan:string \"a plan to execute the task in 5 steps or less\"",
      provider: "google-gemini",
      providerKeyName: "GEMINI_API_KEY",
      ai: {
        model: "gemini-1.5-flash",
        temperature: 0
      },
      options: {
        debug: false
      }
    }
    // ... more agents
  ]
};

// Create a new instance of AxCrew using the config object
const crew = new AxCrew(config);

Both methods support the same configuration structure and options. Choose the one that best fits your use case:

• Use a configuration file when you want to:
  • Keep your configuration separate from your code
  • Share configurations across different projects
  • Version control your configurations
• Use a direct configuration object when you want to:
  • Generate configurations dynamically
  • Modify configurations at runtime
  • Keep everything in one file for simpler projects

Agent Persona: definition and prompt

• definition: Detailed persona/program description used as the system prompt. Recommended for precise control. Must be at least 100 characters.
• prompt: An alias for definition. If both are provided, definition takes precedence.

Add either field to any agent config. The chosen value becomes the Ax agent's underlying program description (system prompt).

{
  "name": "Planner",
  "description": "Creates a plan to complete a task",
  "prompt": "You are a meticulous planning assistant. Produce concise, executable step-by-step plans with clear justifications, constraints, and assumptions. Prefer brevity, avoid fluff, and ask for missing critical details before proceeding when necessary.",
  "signature": "task:string -> plan:string",
  "provider": "google-gemini",
  "providerKeyName": "GEMINI_API_KEY",
  "ai": { "model": "gemini-1.5-flash", "temperature": 0 }
}

Agent Examples

You can provide examples to guide the behavior of your agents using the examples field in the agent configuration. Examples help the agent understand the expected input/output format and improve its responses.

{
  "name": "MathTeacher",
  "description": "Solves math problems with step by step explanations",
  "signature": "problem:string \"a math problem to solve\" -> solution:string \"step by step solution with final answer\"",
  "provider": "google-gemini",
  "providerKeyName": "GEMINI_API_KEY",
  "ai": {
    "model": "gemini-1.5-pro",
    "temperature": 0
  },
  "examples": [
    {
      "problem": "what is the square root of 144?",
      "solution": "Let's solve this step by step:\n1. The square root of a number is a value that, when multiplied by itself, gives the original number\n2. For 144, we need to find a number that when multiplied by itself equals 144\n3. 12 × 12 = 144\nTherefore, the square root of 144 is 12"
    },
    {
      "problem": "what is the cube root of 27?",
      "solution": "Let's solve this step by step:\n1. The cube root of a number is a value that, when multiplied by itself twice, gives the original number\n2. For 27, we need to find a number that when cubed equals 27\n3. 3 × 3 × 3 = 27\nTherefore, the cube root of 27 is 3"
    }
  ]
}

The examples should:

• Match the input/output signature of your agent
• Demonstrate the desired format and style of responses
• Include edge cases or specific patterns you want the agent to learn
• Be clear and concise while showing the expected behavior

Examples are particularly useful for:

• Teaching agents specific response formats
• Demonstrating step-by-step problem-solving approaches
• Showing how to handle edge cases
• Maintaining consistent output styles across responses

Function Registry

Functions (aka Tools) are the building blocks of agents. They are used to perform specific tasks, such as calling external APIs, databases, or other services.

The FunctionRegistry is a central place where all the functions that agents can use are registered. This allows for easy access and management of functions across different agents in the crew.

To use the FunctionRegistry, you need to either:

• import and use the built-in functions from the @amitdeshmukh/ax-crew package, or
• bring your own functions before initializing AxCrew.

Here's an example of how to set up the FunctionRegistry with built-in functions:

import { AxCrewFunctions } from '@amitdeshmukh/ax-crew';
const crew = new AxCrew(configFilePath, AxCrewFunctions);

if you want to bring your own functions, you can do so by creating a new instance of FunctionRegistry and passing it to the AxCrew constructor.

import { FunctionRegistryType } from '@amitdeshmukh/ax-crew';

const myFunctions: FunctionRegistryType = {
  GoogleSearch: googleSearchInstance.toFunction()
};

const crew = new AxCrew(configFilePath, myFunctions);

Adding Agents to the Crew

There are three ways to add agents to your crew, each offering different levels of control:

Method 1: Add All Agents Automatically

This is the simplest method that automatically handles all dependencies:

// Initialize all agents defined in the config
await crew.addAllAgents();

// Get agent instances
const planner = crew.agents?.get("Planner");
const manager = crew.agents?.get("Manager");

This method:

• Reads all agents from your configuration
• Automatically determines the correct initialization order based on dependencies
• Initializes all agents in the proper sequence
• Throws an error if there are circular dependencies

Method 2: Add Multiple Agents with Dependencies

This method allows you to initialize a subset of agents while still handling dependencies automatically:

// Add multiple agents - dependencies will be handled automatically
await crew.addAgentsToCrew(['Manager', 'Planner', 'Calculator']);

// Or add them in multiple steps - order doesn't matter as dependencies are handled
await crew.addAgentsToCrew(['Calculator']); // Will be initialized first
await crew.addAgentsToCrew(['Manager']);    // Will initialize Planner first if it's a dependency

This method:

• Takes an array of agent names you want to initialize
• Automatically handles dependencies even if not explicitly included
• Initializes agents in the correct order regardless of the order specified
• Throws an error if required dependencies are missing or if there are circular dependencies

Method 3: Add Individual Agents

This method gives you the most control but requires manual dependency management:

// Add agents one by one - you must handle dependencies manually
await crew.addAgent('Calculator'); // Add base agent first
await crew.addAgent('Planner');    // Then its dependent
await crew.addAgent('Manager');    // Then agents that depend on both

This method:

• Gives you full control over the initialization process
• Requires you to handle dependencies manually
• Throws an error if you try to initialize an agent before its dependencies

Dependency Handling

The crew system automatically handles agent dependencies in the following ways:

1. Explicit Dependencies: Defined in the agent config using the agents field:

{
  name: "Manager",
  // ... other config ...
  agents: ["Planner", "Calculator"] // Manager depends on these agents
}

2. Initialization Order:

   • Base agents (no dependencies) are initialized first
   • Dependent agents are initialized only after their dependencies
   • Circular dependencies are detected and reported

3. Error Handling:

   • Missing dependencies are reported with clear error messages
   • Circular dependencies are detected and reported
   • Invalid agent names or configurations are caught early

4. State Management:

   • All initialized agents within a crew share the same state
   • Dependencies can access and modify shared state
   • State persists across all initialization methods

Choose the method that best fits your needs:

• Use addAllAgents() for simple cases where you want all agents
• Use addAgentsToCrew() when you need a subset of agents with automatic dependency handling
• Use addAgent() when you need fine-grained control over the initialization process

State Management

The StatefulAxAgent class in src/agents/index.js allows for shared state functionality across agents. Sub-agents can be added to an agent to create complex behaviors. All agents in the crew have access to the shared state. State can also be shared with functions that are passed to the agents. To do this, pass the state object as an argument to the function class as shown here https://axllm.dev/guides/functions-1/

// Set some state (key/value) for this crew
crew.state.set('name', 'Crew1');
crew.state.set('location', 'Earth');

// Get the state for the crew
crew.state.get('name'); // 'Crew1'
crew.state.getAll(); // { name: 'Crew1', location: 'Earth' }

State can also be set/get by individual agents in the crew. This state is shared with all agents. It is also passed to any functions expressed as a class in FunctionsRegistry.

Planner.state.set('plan', 'Fly to Mars'); 
console.log(Manager.state.getAll()); // { name: 'Crew1', location: 'Earth', plan: 'Fly to Mars' }

Example Agent task

An example of how to complete a task using the agents is shown below. The Planner agent is used to plan the task, and the Manager agent is used to execute the task.

import { AxCrew, AxCrewFunctions } from '@amitdeshmukh/ax-crew';

// Create a new instance of AxCrew
const crew = new AxCrew('./agentConfig.json', AxCrewFunctions);
crew.addAgentsToCrew(['Planner', 'Calculator', 'Manager']);

// Get agent instances
const Planner = crew.agents.get("Planner");
const Manager = crew.agents.get("Manager");

// User query
const userQuery = "whats the square root of the number of days between now and Christmas";

console.log(`\n\nQuestion: ${userQuery}`);

// Forward the user query to the agents
const planResponse = await Planner.forward({ task: userQuery });
const managerResponse = await Manager.forward({ question: userQuery, plan: planResponse.plan });

// Get and print the plan and answer from the agents
const plan = planResponse.plan;
const answer = managerResponse.answer;

console.log(`\n\nPlan: ${plan}`);
console.log(`\n\nAnswer: ${answer}`);

Streaming Responses

The package supports streaming responses from agents, allowing you to receive and process agent outputs in real-time. This is particularly useful for long-running tasks or when you want to provide immediate feedback to users.

import { AxCrew, AxCrewFunctions } from '@amitdeshmukh/ax-crew';

// Create and initialize crew as shown above
const crew = new AxCrew('./agentConfig.json', AxCrewFunctions);
await crew.addAgentsToCrew(['Planner']);

const planner = crew.agents.get("Planner");

// Stream responses using the forward method
await planner.forward(
  { task: "Create a detailed plan for a website" },
  {
    onStream: (chunk) => {
      // Process each chunk of the response as it arrives
      console.log('Received chunk:', chunk);
    }
  }
);

// You can also use streaming with sub-agents
await planner.forward(
  ai,
  { task: "Create a detailed plan for a website" },
  {
    onStream: (chunk) => {
      process.stdout.write(chunk);
    }
  }
);

Key streaming features:

• Real-time response processing
• Support for both direct and sub-agent usage
• Customizable stream handling through callbacks
• Compatible with all agent types and configurations
• Maintains cost tracking and state management functionality

Model Context Protocol (MCP) Support

AxCrew provides built-in support for the Model Context Protocol (MCP), allowing agents to connect to and use MCP servers for enhanced functionality. MCP enables agents to access external tools, data sources, and services in a standardized way.

Supported Transport Types

AxCrew supports three MCP transport types, replacing the deprecated AxMCPHTTPTransport:

1. AxMCPStdioTransport - For standard input/output communication
2. AxMCPHTTPSSETransport - For HTTP with Server-Sent Events
3. AxMCPStreambleHTTPTransport - For streamable HTTP communication

Configuration

Add MCP servers to your agent configuration using the mcpServers field:

STDIO Transport Configuration

For MCP servers that communicate via standard input/output:

{
  "name": "DataAnalyst",
  "description": "Analyzes data using MCP tools",
  "signature": "data:string -> analysis:string",
  "provider": "openai",
  "providerKeyName": "OPENAI_API_KEY",
  "ai": {
    "model": "gpt-4",
    "temperature": 0
  },
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/files"],
      "env": {
        "NODE_ENV": "production"
      }
    },
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"]
    }
  }
}

HTTP SSE Transport Configuration

For MCP servers accessible via HTTP with Server-Sent Events:

{
  "name": "WebAnalyst",
  "description": "Analyzes web content using MCP tools",
  "signature": "url:string -> analysis:string",
  "provider": "anthropic",
  "providerKeyName": "ANTHROPIC_API_KEY",
  "ai": {
    "model": "claude-3-haiku",
    "temperature": 0
  },
  "mcpServers": {
    "api-server": {
      "sseUrl": "https://api.example.com/mcp/sse"
    }
  }
}

Streamable HTTP Transport Configuration

For MCP servers that support streamable HTTP communication:

{
  "name": "StreamAnalyst",
  "description": "Processes streaming data using MCP tools",
  "signature": "stream:string -> results:string",
  "provider": "google-gemini",
  "providerKeyName": "GEMINI_API_KEY",
  "ai": {
    "model": "gemini-1.5-pro",
    "temperature": 0
  },
  "mcpServers": {
    "stream-processor": {
      "mcpEndpoint": "http://localhost:3002/stream",
      "options": {
        "authorization": "Bearer ey.JhbGciOiJkaXIiLCJlbmMiOiJBMjU2R0NNIn0..-1234567890.1234567890",
        "headers": { 
          "X-Custom-Header": "custom-value"
        }
      }
    }
  }
}

Mixed Transport Configuration

You can use multiple transport types within the same agent:

{
  "name": "MultiModalAgent",
  "description": "Uses multiple MCP servers with different transports",
  "signature": "task:string -> result:string",
  "provider": "openai",
  "providerKeyName": "OPENAI_API_KEY",
  "ai": {
    "model": "gpt-4",
    "temperature": 0
  },
  "mcpServers": {
    "local-files": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/workspace"]
    },
    "web-search": {
      "sseUrl": "http://localhost:3001/sse"
    },
    "data-stream": {
      "mcpEndpoint": "http://localhost:3002/stream"
    }
  }
}

MCP Server Examples

Here are some popular MCP servers you can use:

Filesystem Server (STDIO):

"filesystem": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-filesystem", "/allowed/path"]
}

Brave Search Server (STDIO):

"brave-search": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-brave-search"],
  "env": {
    "BRAVE_API_KEY": "your-brave-api-key"
  }
}

GitHub Server (STDIO):

"github": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-github"],
  "env": {
    "GITHUB_PERSONAL_ACCESS_TOKEN": "your-github-token"
  }
}

PostgreSQL Server (STDIO):

"postgres": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-postgres"],
  "env": {
    "POSTGRES_CONNECTION_STRING": "postgresql://user:pass@localhost/db"
  }
}

Usage in Code

MCP functions are automatically available to agents once the servers are configured:

import { AxCrew } from '@amitdeshmukh/ax-crew';

// Create crew with MCP-enabled agents
const crew = new AxCrew('./agentConfig.json');
await crew.addAgent('DataAnalyst'); // Agent with MCP servers configured

const analyst = crew.agents.get('DataAnalyst');

// The agent can now use MCP functions automatically
const response = await analyst.forward({
  data: "Please analyze the sales data in /workspace/sales.csv"
});
// The agent will automatically use the filesystem MCP server to read the file
// and any other configured MCP tools for analysis

Best Practices

1. Environment Variables: Store sensitive information like API keys in environment variables rather than in the configuration file.

2. Path Security: For filesystem servers, always specify allowed paths to prevent unauthorized file access.

3. Server Health: Implement health checks for HTTP-based MCP servers to ensure reliability.

4. Error Handling: MCP server failures are handled gracefully - agents will continue to work with available tools.

5. Debugging: Enable debug mode to see MCP server initialization and communication logs:

{
  "debug": true,
  "mcpServers": { ... }
}

Migration from Deprecated Transport

If you're upgrading from the deprecated AxMCPHTTPTransport, update your configuration:

Before (deprecated):

"mcpServers": {
  "my-server": {
    "sseUrl": "http://localhost:3001/sse"
  }
}

After (current): The configuration remains the same - the transport type is automatically detected and AxMCPHTTPSSETransport is used for sseUrl configurations. No changes to your configuration files are needed.

For new streamable HTTP servers, use:

"mcpServers": {
  "my-stream-server": {
    "mcpEndpoint": "http://localhost:3002/stream",
    "options": {
      "timeout": 30000
    }
  }
}

Tracking Usage Costs

The package provides metrics for monitoring API usage across individual agents and the entire crew. Estimated costs are accumulated precisely and rounded to 5 decimal places for reporting.

// After running an agent's forward method
await Planner.forward({ task: userQuery });

// Per-agent metrics
const agentMetrics = (Planner as any).getMetrics?.();
console.log(agentMetrics);
/* Example:
{
  provider: "anthropic",
  model: "claude-3-haiku-20240307",
  requests: { totalRequests, totalErrors, errorRate, totalStreamingRequests, durationMsSum, durationCount },
  tokens: { promptTokens, completionTokens, totalTokens },
  estimatedCostUSD: 0.00091, // rounded to 5 decimals
  functions: { totalFunctionCalls, totalFunctionLatencyMs }
}
*/

// Crew metrics
const crewMetrics = crew.getCrewMetrics();
console.log(crewMetrics);
/* Example:
{
  requests: { ... },
  tokens: { promptTokens, completionTokens, totalTokens },
  estimatedCostUSD: 0.00255, // rounded to 5 decimals
  functions: { ... }
}
*/

// Reset tracked metrics
crew.resetCosts();

Notes:

• Legacy cost APIs (getLastUsageCost, getAccumulatedCosts, getAggregatedCosts) are superseded by metrics methods.
• Estimated cost values are numbers rounded to 5 decimal places.

Changelog

See CHANGELOG.md for a list of changes and version updates.