@computesdk/modal

Modal provider for ComputeSDK — execute code in serverless Modal sandboxes with optional GPU support.

Installation

npm install @computesdk/modal

This package uses the official Modal JS SDK (v0.7).

Setup

Get your Modal token from modal.com
Set the credentials:

export MODAL_TOKEN_ID=your_token_id
export MODAL_TOKEN_SECRET=your_token_secret

Quick Start

import { modal } from '@computesdk/modal';

const compute = modal({
  tokenId: process.env.MODAL_TOKEN_ID,
  tokenSecret: process.env.MODAL_TOKEN_SECRET,
});

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

const result = await sandbox.runCommand(`python - <<'PY'
import sys
print(f"Python {sys.version}")
PY`);

console.log(result.stdout);
await sandbox.destroy();

Configuration

Environment Variables

export MODAL_TOKEN_ID=your_token_id
export MODAL_TOKEN_SECRET=your_token_secret

Configuration Options

interface ModalConfig {
  /** Modal token ID - falls back to MODAL_TOKEN_ID env var */
  tokenId?: string;
  /** Modal token secret - falls back to MODAL_TOKEN_SECRET env var */
  tokenSecret?: string;
  /** Execution timeout in milliseconds */
  timeout?: number;
  /** Modal environment name (e.g. 'main', 'sandbox') */
  environment?: string;
  /** Ports to expose (unencrypted tunnels by default) */
  ports?: number[];
  /** Modal App name (default: 'computesdk-modal') */
  appName?: string;
}

Exposing Ports

Ports declared on the provider are exposed as public Modal tunnels and reachable via getUrl():

const compute = modal({
  tokenId: process.env.MODAL_TOKEN_ID,
  tokenSecret: process.env.MODAL_TOKEN_SECRET,
  ports: [3000, 8080],
});

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

await sandbox.runCommand(`node - <<'JS'
const http = require('http');
http.createServer((req, res) => {
  res.end('Hello from Modal sandbox\\n');
}).listen(3000);
JS`, { background: true });

const url = await sandbox.getUrl({ port: 3000 });
console.log(`Server reachable at: ${url}`);

Note: Tunnels are unencrypted by default for maximum compatibility.

Features

✅ Command Execution — shell commands via Sandbox.exec()
✅ Filesystem Operations — read, write, mkdir, ls, rm
✅ GPU Support — Modal's native GPU access for ML workloads
✅ Serverless Scaling — scale to thousands of concurrent sandboxes
✅ Snapshots — save and restore sandbox state
❌ Interactive Terminals — not exposed by the provider

API Reference

Command Execution

// Run Python code via heredoc
const result = await sandbox.runCommand(`python - <<'PY'
import torch
print(f"CUDA available: {torch.cuda.is_available()}")
PY`);

// Run shell commands directly
await sandbox.runCommand('pip install numpy');
await sandbox.runCommand('ls -la');

// Background process
await sandbox.runCommand('python server.py', { background: true });

// With env vars and cwd
await sandbox.runCommand('python script.py', {
  cwd: '/app',
  env: { DEBUG: 'true' },
});

Filesystem Operations

await sandbox.filesystem.writeFile('/app/script.py', 'print("hello")');
const content = await sandbox.filesystem.readFile('/app/script.py');

await sandbox.filesystem.mkdir('/app/data');
const entries = await sandbox.filesystem.readdir('/app');

const exists = await sandbox.filesystem.exists('/app/script.py');
await sandbox.filesystem.remove('/app/script.py');

Sandbox Management

// Get sandbox info
const info = await sandbox.getInfo();
console.log(info.id, info.provider, info.status);

// Reconnect to existing sandbox
const existing = await compute.instances.getById('sandbox-id');

// Destroy sandbox
await sandbox.destroy();

Error Handling

import { modal } from '@computesdk/modal';

try {
  const compute = modal({
    tokenId: process.env.MODAL_TOKEN_ID,
    tokenSecret: process.env.MODAL_TOKEN_SECRET,
  });
  const sandbox = await compute.instances.create();
  const result = await sandbox.runCommand('python -c "import nonexistent"');

  if (result.exitCode !== 0) {
    console.error('Command failed:', result.stderr);
  }
} catch (error) {
  if (error.message.includes('Missing Modal')) {
    console.error('Set MODAL_TOKEN_ID and MODAL_TOKEN_SECRET');
  } else if (error.message.includes('authentication')) {
    console.error('Check your Modal credentials');
  }
}

Examples

GPU-Accelerated Inference

import { modal } from '@computesdk/modal';

const compute = modal({
  tokenId: process.env.MODAL_TOKEN_ID,
  tokenSecret: process.env.MODAL_TOKEN_SECRET,
});

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

const result = await sandbox.runCommand(`python - <<'PY'
import torch

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")

x = torch.randn(1000, 1000, device=device)
y = torch.matmul(x, x.T)
print(f"Result shape: {y.shape}")
PY`);

console.log(result.stdout);
await sandbox.destroy();

Parallel Task Processing

import { modal } from '@computesdk/modal';

const compute = modal({
  tokenId: process.env.MODAL_TOKEN_ID,
  tokenSecret: process.env.MODAL_TOKEN_SECRET,
});

const tasks = ['task1.json', 'task2.json', 'task3.json'];

const results = await Promise.all(tasks.map(async (taskFile) => {
  const sandbox = await compute.instances.create();
  try {
    return await sandbox.runCommand(`python -c "import json; print(json.dumps({'task': '${taskFile}'}))"`);
  } finally {
    await sandbox.destroy();
  }
}));

console.log(results.map(r => r.stdout));

Best Practices

Resource Management — destroy sandboxes when done; Modal scales but you still pay per use
Error Handling — check exitCode for command failures, catch for SDK/auth errors
GPU Workloads — leverage Modal's GPU offering for ML workloads
Long Tasks — set generous timeout for training or large-file workflows

@computesdk/modal on npm

Recent stable releases

README from npm