@computesdk/aws

AWS providers for ComputeSDK, one product per subpath:

EC2

AWS EC2 provider for ComputeSDK. Each sandbox is one EC2 instance launched from your launch template, on-demand by default or Spot when you opt in. Commands run over SSM Run Command (AWS-RunShellScript). There is no agent or gateway on the box and no inbound ports: the preinstalled SSM agent dials out.

Installation

npm install @computesdk/aws

Usage

import { ec2 } from '@computesdk/aws/ec2';

// On-demand, everything else comes from the launch template
const compute = ec2({ launchTemplate: 'my-template' }); // region from AWS_REGION

const sandbox = await compute.instances.create();
const result = await sandbox.runCommand('node -v', { env: { CI: '1' }, cwd: '/tmp', timeout: 60_000 });
console.log(result.exitCode, result.stdout);
await sandbox.destroy();

Spot

import { ec2, isEc2CapacityError } from '@computesdk/aws/ec2';

const compute = ec2({
  region: 'us-east-2',
  launchTemplate: 'computesdk-market-medium',
  instanceType: [
    'c5d.large', 'm7a.large', 'c6in.large', 'r5a.large', 'm5a.large', 'r5ad.large', 'r5b.large',
    'r5n.large', 'c5a.large', 'r6a.large', 'm5ad.large', 'r6in.large', 'm6in.large',
  ],
  spot: { maxPricePerHour: '0.05' },
});

try {
  const sandbox = await compute.instances.create({ name: 'bench-1', metadata: { listing: 'lst_42' } });
} catch (error) {
  if (isEc2CapacityError(error)) console.log('No Spot capacity right now:', error.codes);
  else throw error;
}

Configuration

Advanced: readyTimeoutMs (default 120000), pollIntervalMs (default 2000), maxOutputBytes (default 4 MiB), outputS3Bucket / outputS3KeyPrefix (see large output below).

The launch template owns the AMI, IMDS settings, root volume, security group, networking, and an instance profile that allows SSM. The provider never sets those.

A plain on-demand launch (at most one instanceType, no subnetIds) is a single RunInstances call. Spot, several instance types, or subnetIds use an instant EC2 Fleet over every type × subnet pair (price-capacity-optimized for Spot).

If your launch template doesn't pin a subnet, pass subnetIds; otherwise AWS falls back to the default VPC.

Credentials

There are no API keys in config. Pass a credential provider for short-lived credentials, e.g. on Vercel with OIDC:

import { awsCredentialsProvider } from '@vercel/functions/oidc';

ec2({ launchTemplate: 'my-template', credentials: awsCredentialsProvider({ roleArn: process.env.AWS_ROLE_ARN! }) });

Behavior

• Sandbox ID is the EC2 instance ID.
• Tags: instances and volumes are tagged computesdk:managed=true (IAM policies can be conditioned on it), computesdk:sandbox-name when name is set, and every string metadata value (sanitized to EC2 tag limits, max 50 tags).
• Capacity: when no instance launches because of capacity, price, or quota errors, create throws Ec2CapacityError (code: 'AWS_EC2_CAPACITY_UNAVAILABLE', codes, errors). Other fleet failures throw Ec2FleetError.
• Readiness: create returns once the instance is Online in SSM. On timeout (Ec2ReadyTimeoutError) or any error the instance is terminated before throwing. create({ timeout }) overrides readyTimeoutMs.
• Commands run as root via AWS-RunShellScript. env is exported and cwd is entered in the same script; timeout maps to SSM executionTimeout (seconds, rounded up). Timed-out commands return exit code 124.
• Large output: SSM returns at most 24,000 characters of stdout and 8,000 of stderr inline. Larger output is written to /tmp/computesdk-ssm/ on the instance and read back in ~17 KB chunks (a few SSM calls in parallel, then deleted), so it arrives complete but slower: roughly 15–30 s per MB. Each stream keeps its last maxOutputBytes (default 4 MiB) and stderr notes any cut. For big output, set outputS3Bucket (the instance profile needs s3:PutObject and the caller s3:GetObject): SSM writes full output there and it is fetched in one request.
• getById returns null for unknown, terminated/shutting-down, or untagged instances. list returns managed pending/running instances. destroy is idempotent and refuses to terminate untagged instances.
• getInstance() / getInfo().metadata expose instance type, AZ, subnet, lifecycle (spot/on-demand), and launch time.
• Not supported: getUrl (no inbound traffic) and the filesystem API.

Testing

pnpm test   # unit tests (mocked AWS)

# real AWS: any Ec2Config as JSON
AWS_EC2_INTEGRATION_CONFIG='{"launchTemplate":"my-template","region":"us-east-2","instanceType":"m7a.large"}' pnpm test

AgentCore

AWS Bedrock AgentCore Code Interpreter provider for ComputeSDK: secure, fully-managed, session-based sandboxes for running code and shell commands, with no infrastructure to provision. Each sandbox maps onto an AgentCore Code Interpreter session.

npm install @computesdk/aws

import { agentcore } from '@computesdk/aws/agentcore';

Previously published as @computesdk/agentcore, which is now a deprecated re-export of this subpath.

Authentication

The provider uses the standard AWS credential provider chain — the same resolution order as the AWS CLI. You don't pass an API key; you authenticate to AWS however you normally would:

• Environment variables (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN)
• An SSO session (aws sso login)
• A named profile (AWS_PROFILE, or profile in config)
• EC2/ECS/EKS instance roles

A region is required, via region in config or AWS_REGION / AWS_DEFAULT_REGION.

IAM permissions

The principal needs these bedrock-agentcore actions on the code interpreter:

• bedrock-agentcore:StartCodeInterpreterSession
• bedrock-agentcore:InvokeCodeInterpreter
• bedrock-agentcore:StopCodeInterpreterSession
• bedrock-agentcore:GetCodeInterpreterSession (for getById)
• bedrock-agentcore:ListCodeInterpreterSessions (for list)

Usage

import { compute } from 'computesdk';
import { agentcore } from '@computesdk/aws/agentcore';

// Uses the default AWS credential chain; region from env or config.
compute.setConfig({ defaultProvider: agentcore({ region: 'us-west-2' }) });

const sandbox = await compute.instances.create();

const result = await sandbox.runCommand('echo "Hello from AgentCore" && python3 -c "print(2 + 2)"');
console.log(result.stdout); // Hello from AgentCore\n4

await sandbox.filesystem.writeFile('/tmp/data.txt', 'persisted across calls');
console.log(await sandbox.filesystem.readFile('/tmp/data.txt'));

await sandbox.destroy();

Using a named profile

agentcore({ region: 'us-west-2', profile: 'my-profile' });

Using explicit / temporary credentials

agentcore({
  region: 'us-west-2',
  credentials: {
    accessKeyId: '...',
    secretAccessKey: '...',
    sessionToken: '...', // temporary credentials supported
  },
});

Configuration

API Reference

Command Execution

// Run a shell command
const result = await sandbox.runCommand('echo "Hello" && ls -la');
console.log(result.stdout, result.stderr, result.exitCode);

// Run Python
const py = await sandbox.runCommand('python3 -c "print(2 + 2)"');

// Run a multi-line script via heredoc
const script = await sandbox.runCommand(`python3 - <<'PY'
import json
print(json.dumps({"ok": True}))
PY`);

// Pass environment variables and a working directory
await sandbox.runCommand('printenv TOKEN', { env: { TOKEN: 'abc' }, cwd: '/tmp' });

// Install packages
await sandbox.runCommand('pip install requests');

runCommand never throws for command failures — a non-zero exit is returned in exitCode, with stdout/stderr captured separately.

Filesystem Operations

await sandbox.filesystem.writeFile('/tmp/hello.py', 'print("Hello World")');
const content = await sandbox.filesystem.readFile('/tmp/hello.py');
await sandbox.filesystem.mkdir('/tmp/data');
const files = await sandbox.filesystem.readdir('/tmp');      // FileEntry[]
const exists = await sandbox.filesystem.exists('/tmp/hello.py');
await sandbox.filesystem.remove('/tmp/hello.py');

Filesystem calls operate on text (UTF-8). To move binary data, fetch it inside the sandbox (runCommand('curl -sL URL -o /tmp/file')) or base64 it over runCommand.

Sandbox Management

const info = await sandbox.getInfo();          // { id, provider, status, createdAt, timeout, metadata }
const existing = await provider.sandbox.getById(sandbox.instanceId);  // or null if terminated
const all = await provider.sandbox.list();     // READY sessions
await sandbox.destroy();

Error Handling

create translates common AWS failures into actionable messages:

try {
  const sandbox = await compute.instances.create();
} catch (err) {
  // e.g. "Missing AWS region for AgentCore. Pass it: agentcore({ region: 'us-west-2' }) ..."
  // or   "AWS authentication failed for AgentCore. Check your credentials ..."
  // or   "Access denied ... Ensure your IAM principal allows bedrock-agentcore:*CodeInterpreter* ..."
  console.error(err.message);
}

• runCommand returns exitCode: 127 with the error on stderr if the invocation itself fails (e.g. the session expired) — it does not throw.
• Filesystem operations throw on failure (e.g. reading a missing file).

Examples

Data processing

const sandbox = await compute.instances.create();

await sandbox.filesystem.writeFile('/tmp/data.csv', 'value\n10\n20\n30\n');
const result = await sandbox.runCommand(`python3 - <<'PY'
import csv, json
with open('/tmp/data.csv') as f:
    values = [int(r['value']) for r in csv.DictReader(f)]
print(json.dumps({'count': len(values), 'sum': sum(values), 'mean': sum(values)/len(values)}))
PY`);
console.log(JSON.parse(result.stdout)); // { count: 3, sum: 60, mean: 20 }

await sandbox.destroy();

As an AI agent tool

The universal InstanceInterface interface maps cleanly onto agent-framework tools — wrap each method in a tool and hand them to your agent (works with any agent SDK):

import { tool } from '@strands-agents/sdk';
import { z } from 'zod';

const runCommand = tool({
  name: 'run_command',
  description: 'Run a shell command in the sandbox.',
  inputSchema: z.object({ command: z.string() }),
  callback: async ({ command }) => {
    const r = await sandbox.runCommand(command);
    return { stdout: r.stdout, stderr: r.stderr, exitCode: r.exitCode };
  },
});
// ...similarly wrap writeFile / readFile / readdir, then: new Agent({ tools: [...] })

Supported runtimes

The managed aws.codeinterpreter.v1 interpreter ships with Python and a standard Linux shell, so any language runtime available in that environment can be driven via runCommand.

Limitations

• No preview URLs / ports. AgentCore Code Interpreter has no inbound network endpoint, so getUrl() throws. Use it for code execution, not for hosting dev servers.
• No interactive PTY. Commands are request/response; there is no bidirectional terminal.
• Sessions expire. A session auto-terminates after its idle timeout; create a new sandbox afterward.
• Filesystem persists, shell environment does not. Files survive across runCommand calls, but each command runs in a fresh shell — cd, export, and shell variables do not carry over. Chain them in a single command (e.g. cd /app && npm test) or use the cwd/env options.
• Background commands don't outlive the call. runCommand(..., { background: true }) returns immediately, but AgentCore terminates the process tree when the invocation ends, so the job is killed rather than left running. Use a single command that runs to completion instead.
• Large writes are chunked. To stay under AgentCore's command-size limit, writeFile uploads base64 in ~60KB chunks — roughly one round-trip per 60KB, so a 5MB file takes ~40s. Reads are single round-trips and fast.
• ~10 concurrent operations per session. AgentCore caps concurrent invocations on a single session; beyond ~10 simultaneous runCommand/filesystem calls, extra ones fail with a quota error. Spread heavy parallelism across multiple sandboxes.

License

MIT