Automate
Browser automation for AI agents, without running the browser.
Your application or agent sends the task and a starting URL. Tabstack handles navigation and interaction inside the call and returns progress events and the final result.
Plain-language tasks · Task events · Guardrails
OPEN
NAVIGATE
READ
RUNNING
- Opened the page
- Chose the date
- Reading the result
- Returning it
The Automate API
Can Tabstack automate a website without my team running a browser?
Yes. Send Tabstack Automate a starting URL and a task. Tabstack runs the browser interaction, streams task events, and returns the completed result without requiring your team to operate browser infrastructure.
hosted automate
navigate, click
fill, paginatepilo, self-hosted
your environment
your controlBrowser automation for AI agents
Your agent describes the task. Tabstack runs the browser.
Your agent writes the task in plain language. Tabstack plans and performs the browser steps and returns the result, so the browser, the page state, and the interaction loop stay outside your model.
An agent calls Automate like any other tool: one call with a task and a starting url, a stream of typed events while the browser works, and the completed result on task:completed. With interactive on, the task can pause and ask for form values.
It plugs into the frameworks you already use: Claude Agent SDK, LangChain, LlamaIndex, Mastra, OpenAI Agents, and Vercel AI SDK.
import Tabstack from '@tabstack/sdk'
const client = new Tabstack()
const stream = await client.agent.automate({
task: 'Name the top 3 trending repositories',
url: 'https://github.com/trending',
})
for await (const event of stream) {
if (event.event === 'task:completed') {
console.log(event.data.finalAnswer)
} else if (event.event === 'error') {
console.error(event.data.error.message)
}
}from tabstack import Tabstack
client = Tabstack()
for event in client.agent.automate(
task='Name the top 3 trending repositories',
url='https://github.com/trending',
):
if event.event == 'task:completed':
print(event.data.final_answer)
elif event.event == 'error':
print(event.data.error.message)Read versus act
Use Automate when reading the page is not enough.
Extract, for content and fields
The answer is already on the page and you need it in a usable shape.
Automate, for interactive flows
The job needs browser actions: navigation, clicks, form completion, pagination, or multi-step interaction.
How it works
How a task runs.
Provide a starting URL
The hosted product begins on the page you name.
Describe the result
State the task you want done in plain language.
Set the controls
Set the action firewall, trusted hosts, and whether it can pause for input.
Stream the progress
Follow task events while the browser is working.
Receive the result
Use the final answer or task output in your application.
The boundary
The hosted Automate endpoint runs tasks on websites.
The hosted Automate endpoint does not log in to private accounts. Do not send account passwords or build a hosted flow that depends on authenticated access. If the automation engine must run under your control, Pilo is the open-source self-hosted path.
Public pages, yes
Public pages that need no sign-in.
Accounts, no
No credentials, and no flow that depends on being signed in.
Your own environment, Pilo
The engine runs where you put it, under your control.
Guardrails
Keep the task inside the actions you allow.
These settings limit what a hosted task can do. For an agent, that is the difference between a task that stays on script and one that page content can redirect. Read how Tabstack handles indirect prompt injection.
| Control | What it governs |
|---|---|
| Action firewall | Which browser actions the task may take |
| Trusted hostnames | Which hosts the task may fill and submit on |
| Human input | When the task pauses for a person |
| Timeouts | How long the task may run |
| Cancellation | How a running task is stopped |
Use cases
Workflows that require interaction.
Submit a search
Fill a search form and return the result.
Compare a catalog
Navigate a catalog and compare the options.
Page through records
Work through records and retrieve the document.
Finish a multi-step flow
Complete a flow and return the final state.
Reach data behind a click
Collect what only appears after an interaction.
Self-hosted
Run the automation engine yourself with Pilo.
Pilo is Mozilla's open-source browser automation engine and the self-hosted path for teams that need control of where the engine runs. The repository carries the deployment instructions and the license.
Read the Pilo introduction andhow interactive mode pauses for a person.
hosted automate
navigate, click
fill, paginatepilo, self-hosted
your environment
your controlChoosing the call
When another option fits better.
- Is the content already on the page? Extract returns it without any interaction.
- Does the task end in a cited answer rather than an action? Researchbuilds the answer from live sources.
- Do you want sessions and browser infrastructure rather than a finished task? A raw browser service is the closer fit.
- Must the engine run in your environment? Pilo is the self-hosted path.
Trust
Know what the hosted task sends and where it runs.
Never trained on. Private by default. Built by Mozilla. The hosted Automate endpoint sends task text and page content to a contracted model provider, and the Trust page lists the account controls.
tabstack
task text and page content
to a contracted providernever trained on
private by default