Skip to main content
Every agent node declares three capability switches. They decide which tools the node gets during an execution. The runtime resolves the node’s toolset from these switches and from the workflow’s environment. The environment wins where the two disagree. The object is optional. Omit it and the runtime applies the defaults below.
Over the API, the SDK, and the MCP tools, the same object sits at properties.capabilities.

​
The three switches

Browser use

Navigate and interact with web pages (browser_use)

Computer use

Vision-based mouse and keyboard control (computer_use)

Ask user question

Pause mid-run and ask a person for input (ask_user_question)

​
Browser use

Gives the node the browser toolset: navigating, clicking, typing, filling forms, reading page content, capturing screenshots, saving PDFs, uploading files, and running Playwright code. Turn it on when the node drives a website. This is how most workflows work on the web. Browser use also opens the node-script bridge that the Google Sheets skill runs on.

​
Computer use

Gives the node vision-based control of a full desktop. The node looks at the screen and moves the mouse and keyboard, instead of working through the DOM. Use it for native applications, and for pages that a DOM-based approach cannot handle. It is slower and costs more than browser use, so prefer browser use when either would work.

​
Ask user question

Lets the node stop and ask a person for a value, a decision, or a confirmation. The execution moves to paused_by_agent until someone answers. Turn it on for workflows with a person in the loop. Leave it off for unattended executions, where a paused run has nobody to answer it.
This switch is separate from email. Inbox tools appear automatically when the execution has a profile with an inbox attached.

Executions and statuses

What paused_by_agent means and how an execution leaves it

​
The environment overrides two of them

Browser use and computer use are resolved against the workflow’s environment.
On a Linux or Windows environment, setting browser_use: true has no effect.If a switch looks like it is being ignored, check the environment first.

Environments

Browser, Linux, and Windows, and what each one runs

​
Runtime features

These features are available to an agent node. Each one becomes usable once its own prerequisite holds.
A feature works only where its prerequisite holds. A workflow that cannot use one says which prerequisite is missing, rather than guessing.

​
Google Sheets

The Google Sheets skill reads and writes cell ranges. It works when all of these are true:
  • browser_use is available, because the skill runs through the browser node-script bridge.
  • The spreadsheet is open in the active browser tab.
  • The profile or session has edit access to the sheet.
  • The target sheet tab is selected.
Google Sheets draws the grid as a canvas. Do not ask the workflow to inspect or edit cells by clicking around the grid. Tell it to use the Google Sheets skill for range reads and writes.

​
Two-factor codes

The Generate TOTP skill produces authenticator codes.
1

Store the secret

Save the TOTP secret on the attached agent profile as a vault credential.
2

Reference it

Write the credential in the node instructions as a ##CREDENTIAL_NAME## placeholder.
3

Time it

Tell the workflow to generate the code immediately before it submits the code.

​
Talking to people and systems

A node has two ways to reach a person:
  • Ask a question. Turn on ask_user_question for nodes that may need an answer mid-run.
  • Send or read email. Attach an agent profile with an inbox. The runtime then exposes send_mail and get_mail on its own.
send_mail and get_mail need a profile attached to the execution, because the inbox address comes from the profile. See Workflow emails for setup and for verification-code patterns.

Nodes

Where capabilities sit on an agent node

Environments

The sandbox that gates browser use and computer use

Agent profiles

Credentials, inboxes, and saved browser state

Workflow emails

Give a workflow an inbox it can send from and read