Extraction Actions

Actions for extracting data and taking screenshots from web pages

Extraction actions retrieve information from web pages using LLM-powered extraction or visual snapshots.

extract

Use an LLM to extract structured or free-text data from the current page’s markdown content.

query · string · required

Description of what data to extract from the page

extract_links · bool · default: "False"

Set to True if the query requires URLs/links, False to save tokens

start_from_char · int · default: "0"

Character position to start extraction from. Use this for long pages when previous extraction was truncated.

Note: This is a character offset in the markdown content, NOT an element index from browser_state.

output_schema · dict

Optional JSON Schema dictionary. When provided, extraction returns validated JSON matching this schema instead of free-text.

See Structured Output Example

Free-text Extraction Example

from browser_use import Agent, Browser, ChatBrowserUse

agent = Agent(
    task="Go to quotes.toscrape.com and extract the first 3 quotes with their authors",
    llm=ChatBrowserUse(),
    browser=Browser()
)

history = await agent.run()
print(history.final_result())

Structured Extraction Example

from browser_use import Agent, Browser, ChatBrowserUse
from pydantic import BaseModel

class Quote(BaseModel):
    text: str
    author: str

class QuotesOutput(BaseModel):
    quotes: list[Quote]

agent = Agent(
    task="Extract the first 5 quotes from quotes.toscrape.com",
    llm=ChatBrowserUse(),
    browser=Browser(),
    output_model_schema=QuotesOutput
)

history = await agent.run()
structured_data = history.structured_output
print(structured_data.quotes)

When to use extract:

• You’re on the right page
• You know what data to extract
• You haven’t called extract before on the same page for the same query

Limitations:

• Cannot extract interactive elements (use browser_state for that)
• Large content may be truncated (use start_from_char to continue)

The extracted content is converted to markdown before extraction, which filters out advertising and noise but removes some interactive elements.

Implementation: browser_use.tools.service:951

---

screenshot

Take a screenshot of the current viewport.

file_name · string

Optional filename to save the screenshot. If provided, saves to file and returns the path.

If omitted, the screenshot will be included in the next browser_state observation.

Supported format: PNG (.png extension added automatically)

Save to File Example

from browser_use import Agent, Browser, ChatBrowserUse

agent = Agent(
    task="Go to example.com and save a screenshot as 'homepage.png'",
    llm=ChatBrowserUse(),
    browser=Browser()
)

await agent.run()

Include in Observation Example

from browser_use import Agent, Browser, ChatBrowserUse

agent = Agent(
    task="Go to example.com and take a screenshot to verify the page loaded",
    llm=ChatBrowserUse(),
    browser=Browser()
)

await agent.run()

Screenshots are useful for:

• Visual confirmation of page state
• Debugging issues
• Creating documentation
• Verifying forms filled correctly

Implementation: browser_use.tools.service:1387

---

search_page

Search page text for a pattern instantly with zero LLM cost (like grep).

pattern · string · required

Text or regex pattern to search for in page content

regex · bool · default: "False"

Treat pattern as regex (default: literal text match)

case_sensitive · bool · default: "False"

Case-sensitive search (default: case-insensitive)

context_chars · int · default: "150"

Characters of surrounding context per match

css_scope · string

CSS selector to limit search scope (e.g., "div#main")

max_results · int · default: "25"

Maximum matches to return

Example

from browser_use import Agent, Browser, ChatBrowserUse

agent = Agent(
    task='Go to wikipedia.org and search for "quantum" on the page',
    llm=ChatBrowserUse(),
    browser=Browser()
)

await agent.run()

Zero LLM cost - This action executes JavaScript directly in the browser for instant results.

Implementation: browser_use.tools.service:1169

---

find_elements

Query DOM elements by CSS selector instantly with zero LLM cost.

selector · string · required

CSS selector to query elements (e.g., "table tr", "a.link", "div.product")

attributes · list[string]

Specific attributes to extract (e.g., ["href", "src", "class"]).

If not set, returns tag and text only.

max_results · int · default: "50"

Maximum elements to return

include_text · bool · default: "True"

Include text content of each element

Example

from browser_use import Agent, Browser, ChatBrowserUse

agent = Agent(
    task='Go to example.com and find all links with their href attributes',
    llm=ChatBrowserUse(),
    browser=Browser()
)

await agent.run()

Use find_elements to:

• Explore page structure
• Count items
• Get links/attributes
• Verify elements exist

Implementation: browser_use.tools.service:1206

---

Related Actions

• Navigation Actions - Navigate between pages
• Interaction Actions - Interact with elements