mainbrella

A Linux computer
for your AI agent.

Start it by API. Run commands, upload files, SSH in, and open web previews. From $5/month.

Start for $5/month Quickstart

Fixed monthly plans. Built on Cloudflare. See limits.

Agent workspaceJavaScript SDK
import { Mainbrella } from '@mainbrella/sdk';
const client = new Mainbrella({
  apiKey: process.env.MAINBRELLA_API_KEY
});
const machine = await client.create({
  catalogId: 'node', size: 'small'
});
try {
  const result = await machine.commands.run(
    'node --version'
  );
  console.log(result.stdout);
} finally { await machine.kill(); }

Give your agent a place to work.

A complete execution workflow, from creation to cleanup.

Explore the API

One account. One workflow.

Named API keys, idempotent creation, and generation checks connect your agent to the same containers you manage in the dashboard. Fixed monthly allowances keep concurrency and session limits explicit.

Private Services, currently a local HTTP prototype, connects machines through named endpoints such as http://api.internal. Attach running machines to an account-owned network; keep the backend private and share the frontend separately.

Isolated from the start.

Cloudflare runs each container in its own microVM with a separate kernel and network. Mainbrella checks ownership, credentials, and container generations at the API boundary.

Security and data handling

Operated by Andrew Arrow, doing business as Mainbrella Co. in Culver City, California. Card details are handled directly by Stripe.

Does Mainbrella fit your workload?

Temporary agent workspaces with a fixed monthly budget.

Compare sandbox and compute platforms
What if I need persistent files?

Save a workspace before stopping to preserve its filesystem and restore it into a fresh container. Ordinary stop discards unsaved changes. Saved workspaces have plan-specific quotas and retention; keep independent exports. See saved workspaces.

Will usage create an unexpected bill?

Builder is $5/month, Pro $180/month, and Scale $999/month, plus applicable taxes. Each includes a fixed compute allowance. Account limits and runtime deadlines cap usage; no automatic overage charges or paid top-ups are enabled. See allowances and a worked example.

Can I use my existing agent?

Use the HTTP API from your application, or follow the Codex, Claude Code, and OpenAI Agents SDK integration recipes. Read the execution and file limits before moving a workload.

Start with a real workload.

Compare all plans →

Builder

$5 / month

Start building
  • 250 compute-unit hours and 1,000 starts/month
  • Five machine sizes, from 256 MiB to 12 GiB RAM
  • Up to five containers within 28 concurrent compute units
  • Up to one-hour sessions · 10-minute idle timeout
  • SSH, browser terminal, files, and web previews

Larger sizes consume more compute units. Hard caps keep usage within your allowance; there is no automatic compute overage billing. Pro and Scale offer larger allowances and longer sessions.

Billed monthly, plus applicable taxes shown at checkout. Terms · Full limits.

Works with your agent.

Use skill instructions with a coding agent, or connect HTTP tools in your application.

Five-step setup instructions

Step 1: Set up Mainbrella

Read https://mainbrella.com/SKILL.md and follow its Quick Setup. It covers the API key handoff (MAINBRELLA_API_KEY, starts with mb_), prerequisites, and the read-only doctor command. Ask the user only where the skill says to.

Step 2: Detect the project

Do not ask. package.json means JavaScript/TypeScript, pyproject.toml or requirements.txt means Python, Cargo.toml means Rust, and go.mod means Go. An empty directory gets a minimal Python project in a new subdirectory unless the user named a language.

Step 3: Integrate

Follow the skill's HTTP workflow, matching the project's package manager, module format, and environment loading. Select an available image for the project. If agent or tool-calling code already exists, wire Mainbrella in where command execution happens and keep the existing conventions.

Step 4: Verify

Run the skill's verification script: create a container, run echo "hello from mainbrella", check stdout and exit code, and stop the created container in a finally block. Pass means stdout "hello from mainbrella", exit code 0, and cleanup completed. If it fails, follow Troubleshooting and re-run once only after cleanup is confirmed and allowance permits. Preserve existing containers.

Step 5: Report

One paragraph: what changed, what ran, and what you verified. Then offer one next step from the skill: a custom image, an LLM tool that runs commands, or a coding agent running inside a container.

Agent setup guide API reference OpenAPI

Already looking for an agent sandbox?

Inspect the open-source platform, or review measured workloads before you start.

Latest updates

Full changelog

— HTTP files, command execution, and idempotent creation

Start with your first workspace.

Sign in and choose a plan to create containers, then connect with SSH or the browser terminal.

Start building — $5/month