@agento-nexus/sdk

Agent OS in Cloud Sandboxes. Run AI agents with full tool access in isolated E2B sandboxes, orchestrate multi-agent workflows, and bridge Claude Code into cloud environments.

Install

npm install @agento-nexus/sdk

Requires Node >= 20. ESM-only.

Core Concepts

FangBox

An isolated cloud sandbox with an OpenFang agent daemon and Claude Code pre-installed. Each FangBox is a full Linux environment where agents can execute code, use tools, and run autonomously.

import { createFangBox } from '@agento-nexus/sdk'

const box = await createFangBox({
  envs: { ANTHROPIC_API_KEY: process.env.ANTHROPIC_API_KEY },
  timeout: 300, // seconds
})

// Deploy an agent using a Hand manifest
await box.client.deployHand({
  name: 'researcher',
  model: 'claude-sonnet-4-6',
  system_prompt: 'You are a research analyst...',
  tools: ['web_search', 'file_read', 'file_write'],
})

// Message the agent
const response = await box.client.message('researcher', {
  role: 'user',
  content: 'Analyze the competitive landscape for AI code assistants',
})

console.log(response.content)

// Clean up
await box.close()

ClaudeBridge

Execute Claude Code prompts directly inside a sandbox — giving Claude full access to the sandbox filesystem, terminal, and tools.

const result = await box.claudeBridge.execute({
  prompt: 'Create a REST API with Express that serves user data from a SQLite database',
  outputFormat: 'json',
  cwd: '/home/user/project',
})

console.log(result.output)  // Claude's response
console.log(result.exitCode) // 0 on success

Fleet

Orchestrate multi-agent workflows as directed acyclic graphs. Each step runs in its own sandbox with dependency resolution and parallel execution.

import { Fleet } from '@agento-nexus/sdk'

const fleet = new Fleet({ maxConcurrency: 3 })

const result = await fleet.run({
  name: 'research-and-report',
  steps: [
    {
      id: 'research',
      hand: { name: 'researcher', system_prompt: '...' },
      prompt: 'Research AI market trends for 2026',
    },
    {
      id: 'analyze',
      hand: { name: 'analyst', system_prompt: '...' },
      prompt: 'Analyze the research findings',
      dependsOn: ['research'],
    },
    {
      id: 'report',
      hand: { name: 'writer', system_prompt: '...' },
      prompt: 'Write an executive report from the analysis',
      dependsOn: ['analyze'],
    },
  ],
})

console.log(result.steps.report.output)

TemplateBuilder

Build custom E2B sandbox templates with pre-installed tools and configurations.

import { TemplateBuilder } from '@agento-nexus/sdk'

const template = new TemplateBuilder({
  base: 'openfang-claude',
  packages: ['postgresql', 'redis'],
  files: { '/etc/openfang/config.toml': configContent },
})

await template.build()

Events

Both FangBox and Fleet emit typed events for monitoring:

box.events.on(event => {
  if (event.type === 'agent:message') {
    console.log(`[${event.agentName}]: ${event.content}`)
  }
})

fleet.events.on(event => {
  if (event.type === 'step:complete') {
    console.log(`Step ${event.stepId} finished in ${event.durationMs}ms`)
  }
})

CLI

npx @agento-nexus/sdk --help

API Reference

Exports

| Export | Description | |--------|-------------| | FangBox | Sandbox wrapper with agent daemon + Claude Code | | createFangBox(config?) | Create and initialize a FangBox | | OpenFangClient | HTTP client for the OpenFang daemon API | | ClaudeBridge | Execute Claude Code inside sandboxes | | Fleet | DAG workflow orchestrator across multiple sandboxes | | TemplateBuilder | Build custom sandbox templates | | retry(fn, opts?) | Retry utility with exponential backoff | | TypedEmitter | Type-safe event emitter | | Logger | Structured logger |

Types

| Type | Description | |------|-------------| | FangBoxConfig | Sandbox creation options | | HandManifest | Agent definition (name, model, system prompt, tools) | | AgentInfo | Running agent status | | AgentMessage | Conversation message | | AgentResponse | Agent reply | | ClaudeCodeRequest | Claude Code execution options | | ClaudeCodeResult | Claude Code execution output | | WorkflowDefinition | Fleet workflow DAG | | WorkflowStep | Single step in a workflow | | StepResult | Output of a completed step | | WorkflowResult | Full workflow execution result |

Requirements

• Node.js >= 20
• E2B API key (E2B_API_KEY)
• Anthropic API key for Claude Code (ANTHROPIC_API_KEY)

License

MIT