# Best sandbox for Mastra agents in 2026

Mastra agents take AI SDK tools as they are, so the best sandbox plugs in as tools and costs almost nothing while idle.

**With Runtime a Mastra agent gets a full Linux microVM from one line,
`tools: runtimeTools(sbx)`, and pays for CPU only while code runs.** The four
tools run commands and read, write and list files in a Firecracker microVM
with its own kernel, where pandas, NumPy and matplotlib are already installed.
Twenty thousand three-minute data questions a month, each busy for 30
CPU-seconds, cost $34.17 on Runtime,
$165.60 on E2B and $165.60 on
Daytona at their published rates, checked 23 September to 2 October 2026.

## How does Mastra connect an agent to a sandbox?

Mastra has two routes. Its workspaces attach a sandbox provider to an agent
with `workspace: new Workspace({ sandbox })`, and Mastra's
[sandbox docs](https://mastra.ai/docs/sandbox/overview) list Daytona, E2B,
Modal, Vercel and Docker among the providers. Separately, any AI SDK tool can
go straight into an agent's `tools`, which is how Runtime plugs in: no
workspace, no provider package, and the same tools work in a plain
`generateText` call.

| Route                                 | Where commands run                  | What you add                        |
| ------------------------------------- | ----------------------------------- | ----------------------------------- |
| `LocalSandbox`, no isolation          | The app's own host                  | Nothing; Mastra calls it not secure |
| `LocalSandbox` with bwrap or seatbelt | The host, with writes limited       | Bubblewrap or macOS Seatbelt        |
| A workspace sandbox provider          | That provider's machines            | The provider's Mastra package       |
| `runtimeTools(sbx)`                   | A Runtime microVM your code created | `withruntime`, one line in `tools`  |

## What matters in a sandbox for Mastra?

- **Per-user machines that cost nothing idle.** A Mastra app usually serves
  many users. `Sandbox.getOrCreate("user-42")` gives each user one named
  sandbox that pauses itself after 60 seconds with nothing happening and
  wakes with its files, memory and processes when that user returns.
- **Packages without a build step.** The default image has Python 3.12,
  Node.js 24, pandas, NumPy and matplotlib, and `pip install` works when the
  agent needs more.
- **Steps that can take a while.** A command runs up to the timeout you give
  it, so a long data job is one tool call, not a polling loop.
- **One tool set across frameworks.** The tools are AI SDK tools, so the same
  sandbox code serves a Mastra agent, an AI SDK route and a workflow step.

## What does a Mastra data agent cost per month?

Twenty thousand questions, each three minutes on a 2 vCPU, 4 GiB sandbox and
busy for 30 CPU-seconds, at each provider's published rates:

| Provider       | Isolation               | CPU billed on     | A month of questions              |
| -------------- | ----------------------- | ----------------- | --------------------------------- |
| Runtime        | Firecracker microVM     | Measured use      | $34.17 |
| Vercel Sandbox | Firecracker microVM     | Measured use      | $106.13  |
| E2B            | Firecracker microVM     | Every vCPU held   | $165.60     |
| Daytona        | Containers by default   | Every vCPU held   | $165.60 |
| Modal          | gVisor, a shared kernel | The CPU requested | $237.96   |

Runtime costs 68% less than Vercel Sandbox and
86% less than Modal for the same month.

## A run on Runtime, recorded 1 October 2026

```ts check
import { Agent } from "@mastra/core/agent";
import { Sandbox } from "withruntime";
import { runtimeTools } from "withruntime/ai";

await using sbx = await Sandbox.create();
const analyst = new Agent({
  id: "analyst",
  name: "Analyst",
  instructions: "Answer with numbers you computed in the sandbox.",
  model: "anthropic/claude-sonnet-4-6",
  tools: runtimeTools(sbx),
});
const answer = await analyst.generate("Total revenue by region in sales.csv?", { maxSteps: 6 });
console.log(answer.text);
```

We ran this agent's `generate()` loop with `@mastra/core` 1.71.0 and the AI
SDK's mock model in place of Claude, so the model's two tool calls were
scripted and Mastra ran them against a live sandbox:

```text
runtime_write_file sales.csv   Wrote 62 bytes to /workspace/sales.csv
runtime_exec python3 -c "...groupby('region').revenue.sum()..."
  exitCode 0
  region
  east     1250
  south     950
  west     1500
```

pandas was there with nothing installed first. Per-user sandboxes and the
rest of the API are in [Mastra in a sandbox](/integrations/mastra).

## When might another sandbox fit better?

- **Mastra's workspace features.** If you want Mastra's workspace layer
  itself, with its filesystem abstraction and mounts, use one of the workspace
  providers Mastra lists; Runtime plugs in as tools instead.

## Sources

Checked 1 October 2026.

- [Mastra sandboxes](https://mastra.ai/docs/sandbox/overview): workspaces,
  the provider list and the `LocalSandbox` isolation levels
- Each provider's published rates, checked 23 September to 2 October 2026, as the
  [pricing guide](/docs/pricing) lists them
- The recorded run: Runtime's TypeScript SDK with `@mastra/core` 1.71.0 and
  `ai` 7.0.87's `MockLanguageModelV4`, against production on 1 October 2026