Using Computer with your agents

Gaëlle Caplier

There are two types of use cases unlocked by Computer: offline and online use cases.

1. Offline use cases

Offline use cases are everything Computer can do without making network requests: script generation, code execution, data analysis, and file manipulation. The agent runs all of it locally inside Computer.

1.1. PPT generation

LLMs produce unreliable results when asked to build slide decks directly. Computer changes that. The agent uses Python libraries like python-pptx to generate a proper .pptx file — with individual slides, embedded charts, and structured text.

The output is a fully editable PowerPoint file you can open, modify, and share.

1.2. File generation

Computer can generate files in formats that are difficult or impossible to produce through prompting alone. PDF generation is a good example: rather than trying to format a document in plain text, the agent uses a Python library to produce a properly structured PDF — with accurate layouts, embedded images, and the right typography.

You specify what you want (format, sections, visual style), and the agent selects the appropriate package and generates the file. The result is a real, downloadable document.

1.3. Document transformation

Any file the agent can read, it can also transform. This includes translating a document to another language, reformatting content, or converting between file types. It works across formats: Word documents, Excel files, PowerPoint presentations, PDFs, Markdown files.

You don't need to specify how the transformation should be done — the agent handles the technical steps. You describe the result you want, and Computer produces it.

1.4. Data analysis on spreadsheets - including large files

Files that are too large to connect as a Dust data source can be uploaded directly to the conversation. Computer mounts them automatically, and the agent can load and process them in full.

For analysis tasks, the agent generates and runs Python code using pre-installed libraries (pandas, NumPy, scikit-learn, seaborn, and others). The output can include charts saved as images, statistical summaries, and a written interpretation of the results. All generated files appear in the conversation and are available to download.

Because the agent both writes and executes the code, it can also explain what it did: which libraries it used, what the charts show, and what the data suggests.

1.5. Reusable and deterministic code

Computer is useful beyond one-off analysis. If you need a workflow that runs the same way every time, the agent can generate the code for it and execute it on demand.

There are two ways to reuse that code:

Ask the agent to generate the code, save it as a file, and upload it to the agent's instructions or skill configuration. The agent will run it each time without regenerating it.
Write or edit the code yourself and upload it directly to the Computer's file system.

This is useful for any recurring task that needs to be deterministic — data formatting, report generation, file transformations — where you want the same result every time without relying on the model to figure it out from scratch.

1.6. Wrap up: an end-to-end workflow with Computer

2. Online use cases

In online mode, the Computer can make outbound network requests. This opens up use cases that depend on live external data, third-party APIs, CLIs, or services that have no MCP integration.

Network access is fully closed by default. All outbound traffic goes through a proxy that denies everything unless a domain has been explicitly approved by a workspace admin

2.1. How the allowlist works

⚠️ Only admins can choose which domains are allowed on your workspace. If you are trying to reach a domain and Computer is failing at that task, that might be because your admins haven't allowed it for your workspace.

There are two layers of domain approval:

Workspace allowlist

Domains approved by a workspace admin that apply to all Computers across the workspace. Any agent using the Computer can reach these domains without additional steps. Admins manage this list in Settings > Computer > Allowed domains.

Agent-requested allowlist

Domains added on the fly during a specific conversation. When Computer needs to reach a domain that isn't on the workspace list, it can request access, and the user approves or rejects it directly in the conversation. Approved domains are scoped to the current Computer session only and do not carry over to other conversations or change workspace-wide settings.

Workspace admins can disable agent-requested domains entirely. When that toggle is off (in Settings > Computer > Allowed domains), Computer can only reach workspace-approved domains and cannot request new ones mid-conversation.

📚 More details on Computer admin setup

2.2. What online mode enables

With a configured allowlist, Computer can:

Call any external API, including APIs with no MCP integration
Authenticate to external services using workspace-stored credentials
Fetch live data from external sources as part of an analysis or workflow
Install additional packages at runtime, provided the relevant package registry is on the domain allowlist (subject to workspace configuration)

Use cases that depend on real-time external data or third-party services fall into this category.

Test Your Knowledge

Answer 5 questions — score 3 or more to pass

Ready to test your understanding of this chapter?

Previous Chapter

← What is Computer?