> ## Documentation Index
> Fetch the complete documentation index at: https://fal.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Sandbox code execution

> fal Agent runs Python in an isolated sandbox for deterministic media work: ffmpeg cuts, pixel-exact edits, format conversion, and video compositing.

Some work must not go through a generative model. A trim, a resize, a frame extraction, a caption overlay, or a format conversion must be exact and repeatable. For this, fal Agent runs Python in an isolated sandbox.

You do not need to ask for code. Say what you want, and the agent decides whether a model or the sandbox is the right tool.

```text theme={null}
Cut the first 2 seconds off the hallway clip, add the caption "Spring drop" at the bottom, and export as a square MP4.
```

## What is available

The sandbox is a Linux environment with Python and common media tools preinstalled:

* **Pillow** and **numpy** for image work
* **ffmpeg** and **ImageMagick** for video, audio, and batch image operations
* **matplotlib** for charts
* **requests** for fetching files
* **Hyperframes**, a renderer that turns HTML and CSS compositions into MP4. The [video-compositing skill](/docs/documentation/agent/skills) governs its use

The agent can install additional Python and system packages inside the sandbox when a task needs them.

## How results come back

Files the agent creates in the sandbox are uploaded to the fal CDN and appear in the chat as generation cells. The cell's pill names the kind of work, such as **video-edit**, **frame-extract**, **image-edit**, or **audio-edit**. Sandbox outputs take an ordinal like any other output. You can pick them from the palette, and they are included in [exports](/docs/documentation/agent/generations/export).

## Video compositing

The **video-compositing** skill covers styled text and graphics over footage: captions and titles, lower thirds, animated logos, kinetic typography, montages, slideshows, and picture-in-picture layouts. The agent writes an HTML composition and renders it to MP4 with Hyperframes inside the sandbox. Pure generation, such as "a video of a dragon", still goes to a video model, and a bare cut still goes to ffmpeg.

## Photo editing

The **photo-editing** skill routes tonal, geometric, and compositing edits to the sandbox. Background removal, brightness and contrast, crop, straighten, and precise compositing are done in code. Edits that change what is in the image, such as "make it winter", go to a generative edit model.

## Limits

| Limit                  | Value                                                                                            |
| :--------------------- | :----------------------------------------------------------------------------------------------- |
| Single command timeout | 300 seconds by default, up to 600                                                                |
| Sandbox work per turn  | About 210 seconds. Longer work continues in the next turn                                        |
| Idle lifetime          | 5 minutes after the last command. Files persist within a chat session while the sandbox is alive |
| Network                | Outbound only. No public ports                                                                   |

Each user gets their own sandbox. Sandboxes are not shared between accounts.

## Privacy

Files enter the sandbox only when the agent downloads them for a task. Results are uploaded to your fal CDN space with the same [retention](/docs/documentation/agent/retention) as any other generation.