@mayurajs/sandbox-agentcore

Amazon Bedrock AgentCore Code Interpreter sessions as sandboxes for mayura/sandbox: a microVM per session, through the official AWS SDK (@aws-sdk/client-bedrock-agentcore).

npm install mayura @mayurajs/sandbox-agentcore

import { createSandboxes } from 'mayura/sandbox';
import { agentCoreSandboxes } from '@mayurajs/sandbox-agentcore';

const sandboxes = createSandboxes(agentCoreSandboxes({ region: 'us-east-1', credentials }), { maxSandboxes: 5, maxLifetimeMs: 3_600_000 });
const sandbox = await sandboxes.create({ lifetimeMs: 900_000 });
const result = await sandbox.exec(['python3', '-c', 'print(6 * 7)']);
await sandbox.release();

• region and credentials (or a function returning them, or a client you own) are options: nothing is read from the environment, AWS config files or instance metadata. Requests go through fetch. IAM needs bedrock-agentcore:StartCodeInterpreterSession, InvokeCodeInterpreter and StopCodeInterpreterSession.
• Sessions start on AWS's own Code Interpreter, aws.codeinterpreter.v1, or the one in codeInterpreter; a sandbox's image names another. A session lives for the sandbox's lifetime (at most 8 hours), and release stops it.
• network states what the Code Interpreter's network mode, set when it was created, lets sessions reach: 'none' (the default) for the sandbox mode, which reaches only S3, or 'all' for the public mode. Mayura cannot change it per session, so sandboxes are created only with that network.
• Commands run through AgentCore's executeCommand: in the background, polled to their end, with output read back in base64 and bounded. AgentCore takes no standard input or environment per command, so both travel as files, and files are written in pieces that fit in a command line (maxCommandBytes, 60,000 by default): large files take many calls. A timeout or cancellation stops every process the command started.
• No ports, desktop, labels, CPUs or memory: sessions have fixed resources.

See the sandbox guide. Apache-2.0.