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

# Compute Sandbox

> Give Airia Agent access to an isolated execution environment for complex, code-driven tasks.

The Compute Sandbox is an isolated execution environment that lets Airia Agent perform complex, code-driven tasks — running custom scripts, using Python libraries, processing files, or executing multi-step logic that goes beyond what a standard tool call can do. When a user with sandbox access submits a request that requires this kind of work, Airia Agent runs the task in a dedicated container. Skills Repositories (collections of scripts and data files you upload) are mounted into that container at execution time, making your custom code and data available to the agent.

## When to use the sandbox

| Scenario                                         | Why the sandbox helps                                                                                                                                                                           |
| ------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Running custom scripts or proprietary algorithms | Execute Python, JavaScript, or shell scripts that aren't exposed as formal tools                                                                                                                |
| Using third-party libraries                      | Install and use packages (e.g., pandas, NumPy, Pillow) that the base agent environment doesn't include                                                                                          |
| Processing large files or datasets               | Operate on data that would exceed an LLM's context window — the sandbox reads, transforms, and summarizes the data before returning a result to the model, so file size isn't a limiting factor |
| Aggregating or joining large datasets            | Perform data joins, aggregations, or filtering across many rows without loading everything into the model's context                                                                             |
| Multi-step data pipelines                        | Chain together extract, transform, and load steps in a single sandboxed run                                                                                                                     |
| Domain-specific configuration or lookup data     | Mount reference files (CSVs, JSONs, config files) via Skills Repositories and access them at execution time                                                                                     |

## Configure sandbox access

<Steps>
  <Step title="Open the Compute Sandbox tab">
    Navigate to **Settings → Airia Agent → Compute Sandbox**.
  </Step>

  <Step title="Enable the sandbox">
    Toggle **Enable Compute Sandbox** on.
  </Step>

  <Step title="Select users or groups">
    Choose which users get sandboxed execution. Options:

    * **All users** — every user with Airia Agent access runs in the sandbox
    * **Specific users and groups** — only selected users run in the sandbox; everyone else uses standard (non-sandboxed) execution
  </Step>

  <Step title="Save">
    Click **Save**. The sandbox is active for the selected users immediately.
  </Step>
</Steps>

<Note>
  Users not assigned to the sandbox continue to use Airia Agent normally — they simply cannot trigger sandboxed execution or benefit from mounted Skills Repositories.
</Note>

## Performance considerations

Sandboxed execution starts a dedicated container for each conversation session. There is a small amount of startup latency the first time a session uses the sandbox (typically a few seconds). Subsequent tasks within the same session reuse the same container and are not significantly slower.

Set expectations with users who are assigned sandbox access — particularly for workflows where they kick off complex tasks immediately after opening Chat.

<Note>
  The compute sandbox becomes significantly more powerful when paired with Skills Repositories — collections of code and data files mounted into the sandbox at execution time. [Manage Skills Repositories →](/admin-hub/airia-agent/skills-repositories)
</Note>