March Agent SDK (TypeScript)

TypeScript SDK for building AI agents in the March AI platform. Provides agent registration, message handling via Kafka (through Agent Gateway), streaming responses, and integrations with LangGraph and Vercel AI SDK.

Quick Start

The fastest way to get started is with the CLI:

npx march-start my-agent
cd my-agent
# Edit .env with your configuration
pnpm dev

Overview

The March Agent SDK (march-ai-sdk) enables developers to build AI agents that integrate with the March AI Management platform. Agents register themselves, receive messages through Kafka topics, and stream responses back to users.

Key Features

• Agent Registration: Automatic registration with AI Inventory
• Message Handling: Function-based handlers with sender filtering
• Response Streaming: Chunked responses via Kafka/Centrifugo
• Conversation History: Access to message history via Conversation Store
• LangGraph Integration: HTTP-based checkpoint saver for LangGraph graphs
• Vercel AI SDK Integration: Persistent message history for Vercel AI SDK
• Attachment Handling: Download images and PDFs for processing
• Long-term Memory: Semantic search across user conversations
• Error Handling: Automatic error responses to users on handler failures

Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│                           Your Agent Application                            │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│    MarchAgentApp                                                            │
│    ├── registerMe() → Agent                                                 │
│    │   ├── onMessage() handler                                              │
│    │   └── streamer() → Streamer                                            │
│    └── run() → Start consume loop                                           │
│                                                                             │
│    Extensions:                                                              │
│    ├── HTTPCheckpointSaver (LangGraph)                                      │
│    └── VercelAIMessageStore (Vercel AI SDK)                                 │
│                                                                             │
└──────────────────────────────────┬──────────────────────────────────────────┘
                                   │ gRPC (Kafka) + HTTP (APIs)
                                   ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                           Agent Gateway                                      │
├─────────────────────────────────────────────────────────────────────────────┤
│  gRPC: AgentStream (bidirectional)      HTTP: /s/{service}/* proxy          │
│  - Auth                                 - /s/ai-inventory/*                 │
│  - Subscribe/Unsubscribe                - /s/conversation-store/*           │
│  - Produce/Consume messages             - /s/ai-memory/*                    │
│                                         - /s/attachment/*                   │
└─────────────────┬───────────────────────────────────┬───────────────────────┘
                  │                                   │
                  ▼                                   ▼
           ┌──────────────┐                  ┌────────────────────┐
           │    Kafka     │                  │   AI Inventory /   │
           │  {agent}.inbox │                │ Conversation Store │
           └──────────────┘                  └────────────────────┘

Message Flow

sequenceDiagram
    participant User
    participant Provider
    participant Router
    participant Gateway as Agent Gateway
    participant Agent as Your Agent
    participant ConvStore as Conversation Store

    User->>Provider: Send message
    Provider->>Router: Kafka: router.inbox
    Router->>Gateway: Kafka: {agent}.inbox
    Gateway->>Agent: gRPC stream: message

    Agent->>ConvStore: Get conversation history
    ConvStore-->>Agent: Message history

    loop Streaming Response
        Agent->>Gateway: gRPC stream: content chunk
        Gateway->>Router: Kafka: router.inbox
        Router->>Provider: Kafka: user.inbox
        Provider->>User: WebSocket: chunk
    end

    Agent->>Gateway: gRPC stream: done=true

Installation

# Using pnpm
pnpm add march-ai-sdk

# Using npm
npm install march-ai-sdk

# Using yarn
yarn add march-ai-sdk

Or scaffold a new project with the CLI:

npx march-start my-agent

Peer Dependencies

The SDK requires the following peer dependencies:

pnpm add @grpc/grpc-js @grpc/proto-loader zod

Optional Extensions

# For LangGraph support
pnpm add @langchain/langgraph @langchain/langgraph-checkpoint @langchain/core

# For Vercel AI SDK support
pnpm add ai @ai-sdk/openai

Quick Start

import { config } from 'dotenv'
import { MarchAgentApp, Message } from 'march-ai-sdk'

// Load environment variables
config({ path: '.env', override: true })

// Initialize the app
const app = new MarchAgentApp({
    gatewayUrl: process.env.GATEWAY_URL || 'agent-gateway:8080',
    apiKey: process.env.GATEWAY_API_KEY || 'your-api-key',
})

async function main() {
    // Register an agent
    const agent = await app.registerMe({
        name: 'my-assistant',
        about: 'A helpful AI assistant',
        document: 'Answers general questions and provides helpful information.',
        representationName: 'My Assistant',
    })

    // Handle messages
    agent.onMessage(async (message: Message, sender: string) => {
        // Access conversation history
        const history = await message.conversation?.getHistory({ limit: 10 })

        // Stream response
        const streamer = agent.streamer(message)
        streamer.stream('Hello! ')
        streamer.stream('How can I help you today?')
        await streamer.finish()
    })

    // Run the agent
    await app.run()
}

main().catch(console.error)

Configuration

Environment Variables

| Variable | Description | Default | |----------|-------------|---------| | GATEWAY_URL | Agent Gateway endpoint | Required | | GATEWAY_API_KEY | API key for authentication | Required | | HEARTBEAT_INTERVAL | Heartbeat frequency (seconds) | 60 | | CONNECTION_SECURE | Use TLS for connections | false |

MarchAgentApp Options

interface AppOptions {
    gatewayUrl: string       // Gateway endpoint (host:port)
    apiKey: string           // Authentication key
    heartbeatInterval?: number  // Heartbeat frequency in seconds (default: 60)
    maxConcurrentTasks?: number // Max concurrent message handlers (default: 100)
    errorMessageTemplate?: string // Error message sent to users
    secure?: boolean         // Use TLS for gRPC and HTTPS for HTTP (default: false)
}

const app = new MarchAgentApp({
    gatewayUrl: 'agent-gateway:8080',
    apiKey: 'your-api-key',
    heartbeatInterval: 60,
    secure: false,
    errorMessageTemplate: 'Sorry, something went wrong. Please try again.',
})

API Reference

MarchAgentApp

Main application class for initializing and running agents.

import { MarchAgentApp } from 'march-ai-sdk'

const app = new MarchAgentApp({
    gatewayUrl: 'agent-gateway:8080',
    apiKey: 'your-api-key',
})

// Register an agent
const agent = await app.registerMe({
    name: 'agent-name',           // Unique identifier (used for routing)
    about: 'Short description',   // Brief description for agent selection
    document: 'Full docs...',     // Detailed documentation
    representationName: 'Name',   // Display name (optional)
    baseUrl: 'http://...',        // Base URL for artifacts (optional)
    metadata: { key: 'value' },   // Additional metadata (optional)
    relatedPages: [               // Related pages (optional)
        { name: 'Dashboard', endpoint: '/dashboard' }
    ],
})

// Start consuming messages (blocks until shutdown)
await app.run()

Agent

Handles message registration and response streaming.

// Register message handler (all senders)
agent.onMessage(async (message, sender) => {
    // Handle message
})

// Filter by sender
agent.onMessage(async (message, sender) => {
    // Only handles messages from 'user'
}, { senders: ['user'] })

// Exclude sender
agent.onMessage(async (message, sender) => {
    // Handles all except 'other-agent'
}, { senders: ['~other-agent'] })

Message

Represents an incoming message with conversation context.

agent.onMessage(async (message: Message, sender: string) => {
    // Core properties
    message.id                  // Message ID
    message.conversationId      // Conversation ID
    message.userId              // User who sent the message
    message.content             // Message content
    message.createdAt           // Timestamp

    // Metadata
    message.metadata            // Custom metadata from frontend
    message.schema              // JSON schema for form responses

    // Attachments
    message.hasAttachment()     // Check if attachment exists
    message.attachment          // AttachmentInfo object
    await message.getAttachmentBytes()   // Download as Buffer
    await message.getAttachmentBase64()  // Download as base64 string

    // Conversation history
    await message.conversation?.getHistory({ limit: 10 })
    await message.conversation?.getAgentHistory({ limit: 10 })

    // Long-term memory
    await message.memory?.queryAboutUser({ query: '...', limit: 5 })
    await message.memory?.getUserSummary()
})

Message Properties

| Property | Type | Description | |----------|------|-------------| | id | string | Message ID | | conversationId | string | Conversation ID | | userId | string | User who sent the message | | content | string | Message content | | createdAt | string | ISO timestamp | | metadata | Record<string, unknown> \| undefined | Custom metadata | | schema | Record<string, unknown> \| undefined | JSON schema for forms | | attachment | AttachmentInfo \| undefined | File attachment metadata | | conversation | Conversation \| undefined | Conversation helper | | memory | Memory \| undefined | Long-term memory helper |

Streamer

Streams response chunks back to the user (or another agent).

agent.onMessage(async (message, sender) => {
    // Basic streaming
    const streamer = agent.streamer(message)
    streamer.stream('Chunk 1...')
    streamer.stream('Chunk 2...')
    await streamer.finish()

    // Streaming with options
    const streamer2 = agent.streamer(message, {
        sendTo: 'user',      // or another agent name
        awaiting: true,      // Set awaiting_route to this agent
    })

    // Non-persisted content (not saved to DB)
    streamer2.stream('Thinking...', { persist: false, eventType: 'thinking' })

    // Persisted content
    streamer2.stream('Here is my response...')

    // Set response schema for forms
    streamer2.setResponseSchema({
        type: 'object',
        properties: {
            name: { type: 'string' }
        }
    })

    // Set message metadata
    streamer2.setMessageMetadata({
        model: 'gpt-4o',
        confidence: 0.95,
    })

    await streamer2.finish()
})

Streamer Methods

| Method | Description | |--------|-------------| | stream(content, options?) | Stream a content chunk | | write(content, persist?) | Alias for stream() | | finish(awaiting?) | Finish streaming with done=true signal | | setResponseSchema(schema) | Set JSON Schema for form rendering | | setMessageMetadata(metadata) | Set message metadata | | streamBy(structural) | Bind a structural streamer for artifacts, surfaces, etc. |

Structural Streaming

The SDK provides structural streaming components for rich content like artifacts, surfaces, text blocks, and steppers.

Artifact

Stream file artifacts (images, documents, iframes) with progress indication:

import { Artifact } from 'march-ai-sdk'

agent.onMessage(async (message, sender) => {
    const streamer = agent.streamer(message)
    
    // Create an artifact
    const artifact = new Artifact()
    
    // Signal generation started (optional)
    streamer.streamBy(artifact).generating('Creating chart...', 0.5)
    
    // ... do work ...
    
    // Signal completion with artifact details
    streamer.streamBy(artifact).done({
        url: 'https://example.com/chart.png',
        type: 'image',
        title: 'Sales Chart',
        description: 'Q4 2024 sales data',
        metadata: { size: 1024, mimeType: 'image/png' }
    })
    
    streamer.stream('Here is your chart!')
    await streamer.finish()
})

Surface

Stream embedded interactive components (iframes, embeds):

import { Surface } from 'march-ai-sdk'

const surface = new Surface()
streamer.streamBy(surface).generating('Loading calendar...')
streamer.streamBy(surface).done({
    url: 'https://cal.com/embed/user',
    type: 'iframe',  // default
    title: 'Schedule a Meeting'
})

TextBlock

Stream collapsible text content with titles and variants:

import { TextBlock } from 'march-ai-sdk'

const block = new TextBlock()
streamer.streamBy(block).setVariant('thinking')
streamer.streamBy(block).streamTitle('Analyzing...')
streamer.streamBy(block).streamBody('Step 1: Checking data\n')
streamer.streamBy(block).streamBody('Step 2: Processing\n')
streamer.streamBy(block).updateTitle('Analysis Complete')
streamer.streamBy(block).done()

Variants: thinking, note, warning, error, success

Stepper

Stream multi-step progress indicators:

import { Stepper } from 'march-ai-sdk'

const stepper = new Stepper({ steps: ['Fetch Data', 'Process', 'Generate Report'] })
streamer.streamBy(stepper).startStep(0)
// ... do work ...
streamer.streamBy(stepper).completeStep(0)
streamer.streamBy(stepper).startStep(1)
streamer.streamBy(stepper).addStep('Verify')  // Add dynamic step
streamer.streamBy(stepper).completeStep(1)
streamer.streamBy(stepper).startStep(2)
streamer.streamBy(stepper).completeStep(2)
streamer.streamBy(stepper).done()

Structural Streaming Components

| Component | Purpose | Key Methods | |-----------|---------|-------------| | Artifact | Files, images, documents | generating(), done(), error() | | Surface | Embedded iframes/widgets | generating(), done(), error() | | TextBlock | Collapsible text content | streamTitle(), streamBody(), setVariant(), done() | | Stepper | Multi-step progress | startStep(), completeStep(), failStep(), addStep(), done() |

Extensions

LangGraph Integration

Use HTTPCheckpointSaver to persist LangGraph state via the Conversation Store checkpoint API.

import { MarchAgentApp, Message } from 'march-ai-sdk'
import { HTTPCheckpointSaver } from 'march-ai-sdk/extensions/langgraph'
import { StateGraph, START, END, Annotation } from '@langchain/langgraph'
import { ChatOpenAI } from '@langchain/openai'
import { HumanMessage, AIMessage, BaseMessage } from '@langchain/core/messages'

const app = new MarchAgentApp({
    gatewayUrl: process.env.GATEWAY_URL!,
    apiKey: process.env.GATEWAY_API_KEY!,
})

// Create checkpointer
const checkpointer = new HTTPCheckpointSaver(app)

// Define state
const StateAnnotation = Annotation.Root({
    messages: Annotation<BaseMessage[]>({
        reducer: (x, y) => x.concat(y),
    }),
})

// Create graph
const llm = new ChatOpenAI({ model: 'gpt-4o-mini', streaming: true })

async function respond(state: typeof StateAnnotation.State) {
    const response = await llm.invoke(state.messages)
    return { messages: [response] }
}

const graph = new StateGraph(StateAnnotation)
    .addNode('respond', respond)
    .addEdge(START, 'respond')
    .addEdge('respond', END)

const compiled = graph.compile({ checkpointer })

// Register agent
const agent = await app.registerMe({
    name: 'langgraph-agent',
    about: 'Agent using LangGraph for state management',
    document: 'Uses LangGraph to maintain conversation state.',
})

agent.onMessage(async (message: Message, sender: string) => {
    const config = { configurable: { thread_id: message.conversationId } }

    // Add user message and invoke graph
    const result = await compiled.invoke(
        { messages: [new HumanMessage(message.content)] },
        config
    )

    // Stream response
    const streamer = agent.streamer(message)
    const lastMessage = result.messages[result.messages.length - 1]
    streamer.stream(lastMessage.content as string)
    await streamer.finish()
})

await app.run()

Vercel AI SDK Integration

Use VercelAIMessageStore to persist message history for Vercel AI SDK.

import { MarchAgentApp, Message } from 'march-ai-sdk'
import { VercelAIMessageStore } from 'march-ai-sdk/extensions/vercel-ai'
import { streamText, CoreMessage } from 'ai'
import { openai } from '@ai-sdk/openai'

const app = new MarchAgentApp({
    gatewayUrl: process.env.GATEWAY_URL!,
    apiKey: process.env.GATEWAY_API_KEY!,
})

// Create message store
const store = new VercelAIMessageStore(app)

const agent = await app.registerMe({
    name: 'vercel-ai-agent',
    about: 'Agent using Vercel AI SDK',
    document: 'Uses Vercel AI SDK for LLM interactions.',
})

agent.onMessage(async (message: Message, sender: string) => {
    // Load previous messages
    const history = await store.load(message.conversationId)

    // Build messages array
    const messages: CoreMessage[] = [
        { role: 'system', content: 'You are a helpful assistant.' },
        ...history,
        { role: 'user', content: message.content }
    ]

    const streamer = agent.streamer(message)

    // Stream response
    const result = await streamText({
        model: openai('gpt-4o'),
        messages,
        onChunk: ({ chunk }) => {
            if (chunk.type === 'text-delta') {
                streamer.stream(chunk.textDelta)
            }
        }
    })

    await streamer.finish()

    // Save updated history
    await store.save(message.conversationId, [
        ...history,
        { role: 'user', content: message.content },
        { role: 'assistant', content: result.text }
    ])
})

await app.run()

VercelAIMessageStore Methods

| Method | Description | |--------|-------------| | load(conversationId) | Load message history | | save(conversationId, messages) | Save message history | | clear(conversationId) | Clear message history | | append(conversationId, messages) | Append to existing history | | getLastMessages(conversationId, count) | Get last N messages |

Advanced Usage

Inter-Agent Communication

Agents can send messages to other agents using the sendTo option.

// Handle user messages - forward to specialist
agent.onMessage(async (message, sender) => {
    const streamer = agent.streamer(message, { sendTo: 'specialist-agent' })
    streamer.stream('Forwarding your question to the specialist...')
    await streamer.finish()
}, { senders: ['user'] })

// Handle specialist responses - send back to user
agent.onMessage(async (message, sender) => {
    const streamer = agent.streamer(message, { sendTo: 'user' })
    streamer.stream(`The specialist says: ${message.content}`)
    await streamer.finish()
}, { senders: ['specialist-agent'] })

Dynamic Forms with Response Schema

Request structured input from users using JSON Schema forms.

agent.onMessage(async (message, sender) => {
    // Check if this is a form response
    if (message.schema) {
        const data = JSON.parse(message.content)
        const streamer = agent.streamer(message)
        streamer.stream(`Thank you, ${data.name}! I received your message.`)
        await streamer.finish()
        return
    }

    // Request form input
    const streamer = agent.streamer(message, { awaiting: true })
    streamer.setResponseSchema({
        type: 'object',
        title: 'Contact Information',
        properties: {
            name: { type: 'string', title: 'Full Name' },
            email: { type: 'string', format: 'email', title: 'Email' },
            message: { type: 'string', title: 'Message' }
        },
        required: ['name', 'email']
    })
    streamer.stream('Please fill out your contact information:')
    await streamer.finish()
})

Error Handling

Customize error handling behavior:

const app = new MarchAgentApp({
    gatewayUrl: '...',
    apiKey: '...',
    errorMessageTemplate: 'Sorry, something went wrong. Please try again.',
})

const agent = await app.registerMe({ ... })

// Disable automatic error responses
agent.sendErrorResponses = false

agent.onMessage(async (message, sender) => {
    try {
        // Your logic
    } catch (error) {
        // Custom error handling
        const streamer = agent.streamer(message)
        streamer.stream(`I encountered an issue: ${error}`)
        await streamer.finish()
    }
})

Project Structure

ai-ts-framework/
├── src/
│   ├── index.ts              # Package exports
│   ├── app.ts                # MarchAgentApp
│   ├── agent.ts              # Agent class
│   ├── message.ts            # Message class
│   ├── streamer.ts           # Streamer class
│   ├── conversation.ts       # Conversation helper
│   ├── memory.ts             # Memory helper
│   ├── conversation-client.ts # HTTP client for Conversation Store
│   ├── checkpoint-client.ts  # HTTP client for checkpoints
│   ├── agent-state-client.ts # HTTP client for agent state
│   ├── memory-client.ts      # HTTP client for AI Memory
│   ├── attachment-client.ts  # HTTP client for attachments
│   ├── gateway-client.ts     # gRPC client for Agent Gateway
│   ├── heartbeat.ts          # Heartbeat manager
│   ├── api-paths.ts          # API endpoint configuration
│   ├── types.ts              # TypeScript types
│   ├── exceptions.ts         # Custom exceptions
│   ├── structural/           # Structural streaming components
│   │   ├── index.ts          # Structural exports
│   │   ├── base.ts           # StructuralStreamer base class
│   │   ├── artifact.ts       # Artifact component
│   │   ├── surface.ts        # Surface component
│   │   ├── text-block.ts     # TextBlock component
│   │   └── stepper.ts        # Stepper component
│   ├── proto/
│   │   └── gateway.proto     # gRPC protocol definition
│   └── extensions/
│       ├── index.ts
│       ├── langgraph.ts      # LangGraph HTTPCheckpointSaver
│       └── vercel-ai.ts      # Vercel AI MessageStore
├── tests/
│   ├── agent.test.ts
│   ├── app.test.ts
│   ├── checkpoint-client.test.ts
│   ├── message.test.ts
│   ├── streamer.test.ts
│   └── structural-streaming.test.ts
├── tsup.config.ts            # Build configuration
├── vitest.config.ts          # Test configuration
├── tsconfig.json
├── package.json
└── README.md

Development

Running Tests

# Install dependencies
pnpm install

# Run tests
pnpm test

# Run tests in watch mode
pnpm test:watch

# Run tests with coverage
pnpm test:coverage

Building

# Build the package
pnpm build

# Type check without building
pnpm typecheck

Dependencies

| Package | Purpose | |---------|---------| | @grpc/grpc-js | gRPC communication with Agent Gateway | | @grpc/proto-loader | Protobuf loading | | zod | Schema validation |

Optional Dependencies

| Package | Install Command | Purpose | |---------|-----------------|---------| | @langchain/langgraph | pnpm add @langchain/langgraph | LangGraph state management | | @langchain/langgraph-checkpoint | pnpm add @langchain/langgraph-checkpoint | Checkpoint interface | | ai | pnpm add ai | Vercel AI SDK | | @ai-sdk/openai | pnpm add @ai-sdk/openai | OpenAI provider for Vercel AI |

TypeScript Types

The SDK exports all types for TypeScript users:

import type {
    // Core types
    Message,
    Streamer,
    Agent,
    MarchAgentApp,

    // Structural streaming
    StructuralStreamer,
    Artifact,
    Surface,
    TextBlock,
    Stepper,

    // Options
    AppOptions,
    RegisterOptions,
    StreamOptions,
    StreamerOptions,

    // Data types
    KafkaMessage,
    AgentRegistrationData,
    AttachmentInfo,
    ConversationData,

    // Memory types
    MemoryMessage,
    MemorySearchResult,
    UserSummary,

    // Handler types
    MessageHandler,
    SenderFilterOptions,
} from 'march-ai-sdk'

Exceptions

The SDK provides typed exceptions:

import {
    MarchAgentError,      // Base error class
    RegistrationError,    // Agent registration failed
    KafkaError,          // Kafka communication error
    ConfigurationError,  // Invalid configuration
    APIException,        // HTTP API error
    HeartbeatError,      // Heartbeat failed
    GatewayError,        // Gateway communication error
} from 'march-ai-sdk'

License

MIT