Introduction

​

Overview

In a world where AI agents can reason, plan, and write code, AVM provides the missing infrastructure to let them act.

AVM (Agent Virtual Machine) is the runtime layer for the agent economy: a secure, scalable compute protocol that turns agent outputs into real-world execution. It bridges large language models (LLMs) and code execution, allowing agents to autonomously run scripts, process data, and interact with systems without infrastructure overhead.

At the core of AVM is the Model Context Protocol (MCP) a lightweight, JSON-based standard for exposing tools to LLMs. AVM runs an MCP server that accepts tool calls from agents (e.g., run_code), executes them in isolated containers, and returns verifiable outputs, all in real-time, and without relying on centralized APIs.

​

Core Advantages

• One Standard Interface
  Works out-of-the-box with Claude, GPT-4, Cursor, and any agent using MCP or function calling.
• Secure, Auditable Execution
  Every task runs in a sandboxed container, with outputs like stdout, stderr, and exit codes returned to the agent.
• Zero DevOps Overhead
  No servers, no scaling logic just hit the API or SDK and start executing.
• Enterprise-Grade Performance
  Autoscaled, containerized runtimes optimized for ML, cryptography, and agent workloads.
• Flexible Pricing
  Pay-per-run with support for fiat, crypto, and X402 payment rails.

​

Architecture

The AVM execution flow is simple yet powerful:

1. Agent generates code through your LLM
2. MCP Server routes the execution job
3. AVM Node executes code in isolated sandbox
4. Returns structured output: stdout, stderr, exit code, and metadata

​

Get Started

1. Generate an API Key
   Visit the API Keys page and create a new key to authenticate Sandbox Runs API calls.
2. Configure Your Agent
   Point your MCP client to the AVM server:

   {
     "mcpServers": {
       "avm": {
         "command": "npx",
         "args": ["@avm-ai/avm-mcp"],
         "env": { "AVM_API_KEY": "YOUR_KEY_HERE" }
       }
     }
   }
   

3. Launch Jobs
   Send your first code task via the Runs API or through your LLM’s toolchain and watch AVM return structured outputs.