Anvia Docs

Use createSandboxTools(...) when an agent should run commands and manage files inside a live sandbox session.

import { AgentBuilder } from "@anvia/core/agent";
import { createSandboxTools, DockerSandbox } from "@anvia/sandbox";

const sandbox = new DockerSandbox();
const session = await sandbox.createSession();

const tools = createSandboxTools(session, {
  allow: ["exec_command", "read_file", "write_file", "list_files"],
  exec: {
    allowedCommands: ["node", "npm"],
    defaultTimeoutMs: 10_000,
    maxTimeoutMs: 30_000,
  },
  readFile: {
    maxBytes: 64_000,
  },
  writeFile: {
    maxBytes: 64_000,
  },
});

const agent = new AgentBuilder("coder", model)
  .instructions("Use the sandbox for code execution and file operations.")
  .tools(tools)
  .defaultMaxTurns(6)
  .build();

try {
  const response = await agent
    .prompt("Write a JavaScript file that prints the first five Fibonacci numbers, run it, and summarize the output.")
    .send();

  console.log(response.output);
} finally {
  await session.destroy();
}

The default bundle exposes:

Tool	Purpose
`exec_command`	Run a structured command with `command` and `args`
`read_file`	Read a text file from the sandbox workspace
`write_file`	Write a text file, creating parent directories
`list_files`	List files and directories under a workspace path

Keep the session lifetime explicit. Create the session before the agent run and destroy it in a finally block.

Tool Policy

Use allow to choose which sandbox tools are exposed. Use exec.allowedCommands or exec.blockedCommands to keep model-chosen commands inside the product boundary.

const tools = createSandboxTools(session, {
  allow: ["exec_command", "read_file", "list_files"],
  exec: {
    allowedCommands: ["node", "npm"],
    maxTimeoutMs: 30_000,
  },
  readFile: {
    maxBytes: 64_000,
  },
});

Tool policy is separate from Docker isolation. Docker limits protect the host; tool policy controls what the model can ask the session to do.

Agent Tools

Tool Policy