Block Types and Configuration

Reference guide for every agent block type in the Skyvern Cloud UI, including Browser Task, Extraction, Login, Loop, While Loop, Conditional, Code, HTTP Request, and more.

Agents are built from blocks. Each block performs one specific operation: navigating a page, extracting data, making an API call, or branching logic. This page covers every block type available in the Cloud UI agent editor, grouped by category.

If you’re writing automations in code instead, the equivalent operations are Page and Agent methods. See the Actions Reference for the code-first counterparts to these blocks.

To add a block in your agent, click the + button, click on Add block, then select the block type from the menu.

Quick reference

Category	Blocks
Browser Automation	Browser Task, Browser Action, Extraction, Login, Go to URL, Print Page
Data and Extraction	Web Search, Text Prompt, File Parser
Control Flow	Loop, While Loop, Conditional, AI Validation, Code, Wait
Files	File Download, Cloud Storage Upload
Communication	Send Email, HTTP Request, Human Interaction

---

Common fields

These fields appear on most blocks:

Field	Description
Label	Display name on the canvas. Also determines how downstream blocks reference this block’s output ({{ label_output }}).
Continue on Failure	Allow the agent to continue if this block fails
Next Loop on Failure	Only inside loops. Skip to the next iteration on failure.
Model	Override the default LLM model for this block (Advanced Settings on most blocks)

Browser-based blocks (Browser Task, Browser Action, Extraction, Login, File Download) share these additional fields:

Field	Description
URL	Starting URL. Leave blank to continue from the previous block’s page.
Engine	AI engine. Skyvern 1.0 is the default for new browser blocks.
Max Steps Override	Cap the number of AI reasoning steps
Disable Cache	Skip cached script execution
TOTP Identifier	Link TOTP credentials for 2FA handling
TOTP Verification URL	Endpoint this block polls when it needs a 2FA code
Error Code Mapping	Map error conditions to custom error codes (JSON)

Set TOTP Verification URL on the block that may encounter 2FA. Top-level workflow run totp_url values do not automatically populate empty block fields.

---

Browser Automation

Browser Task

The primary block for browser automation. Accepts a natural-language prompt and autonomously navigates the browser to accomplish the goal.

The block uses Skyvern 1.0 by default. Existing agents that already contain V2 browser blocks remain editable and may show a reduced field set.

Skyvern 1.0:

Field	Type	Description
URL	text	Starting URL (optional)
Prompt	text	Natural language instructions for the AI

Existing V2 blocks:

Field	Type	Description
URL	text	Starting URL (optional)
Prompt	text	Natural language instructions for the AI
Max Steps	number	Maximum AI steps
Disable Cache	boolean	Skip cached script execution

Additional fields in Advanced Settings (Skyvern 1.0 only):

Field	Type	Description
Complete if…	text	Condition that signals the block is done
Include Action History in Verification	boolean	Include prior actions when verifying completion
Complete on Download	switch	Mark the block as complete when a file download occurs
File Name	text	Custom filename for downloaded files

Plus common browser fields.

Browser Task is the recommended block for most browser automation. Use it for anything from form filling to multi-page navigation. Include your success criteria directly in the prompt so the AI knows when it’s done.

Browser Action

Execute a single browser action. Best for precise, one-step operations like clicking a specific button or entering text in a field.

Field	Type	Description
URL	text	Starting URL (optional)
Action Instruction	text	Specify the single action to perform

Additional fields in Advanced Settings:

Field	Type	Description
Complete on Download	switch	Mark the block as complete when a file download occurs
File Name	text	Custom filename for downloaded files

Plus common browser fields.

Extraction

Extract structured data from the current page using AI.

Field	Type	Description
Data Extraction Goal	text	What data to extract
Data Schema	JSON	JSON Schema defining the output format. Enable the checkbox to reveal the editor. Click Generate with AI to auto-suggest a schema from your extraction goal.

Plus common browser fields.

Login

Authenticate to a website using stored credentials. Pre-filled with a default login goal that handles common login flows including 2FA.

Field	Type	Description
URL	text	Login page URL
Login Goal	text	Login instructions (pre-filled with a comprehensive default)
Credential	selector	Select stored credentials for authentication

Additional fields in Advanced Settings:

Field	Type	Description
Complete if…	text	How to know login succeeded

Plus common browser fields.

Go to URL

Navigate the browser directly to a specific URL without AI interaction.

Field	Type	Description
URL	text	(Required) The URL to navigate to. Supports parameter references.

Print Page

Print the current browser page to a PDF file. The PDF is saved as a downloadable artifact.

Field	Type	Description
Page Format	select	Paper size: A4, Letter, Legal, or Tabloid
Print Background	boolean	Include CSS background colors and images in the PDF
Headers & Footers	boolean	Add date, title, URL, and page numbers to the PDF

Additional fields in Advanced Settings:

Field	Type	Description
Custom Filename	text	Custom name for the generated PDF
Landscape	boolean	Use landscape orientation

---

Data and Extraction

Web Search

Search Google or Exa over an API. Add site:example.com to the query to search within a site. This block does not open a browser or read linked pages.

Google supports site restrictions. With a single site:domain or site:domain/path filter, the block removes results outside the filter. It repeats the search up to twice when Google returns only results outside that filter. It does not filter queries with quotation marks, parentheses, |, OR, or multiple site: filters. Exa supports one positive site:domain filter. Unsupported Exa expressions fail instead of expanding the search.

Field	Type	Description
Search Query	text	Required query. Supports input and output parameter references.
Search Provider	selector	Automatic (default), Google, or Exa. Google uses SerpAPI.
Maximum Results	integer	Between 1 and 100; default 10. The provider may return fewer results.
Prompt	text	Optional instructions to summarize, rank, extract, or transform results.
Model	selector	Model for Prompt, Data Schema, and error detection.
Data Schema	JSON	Optional output shape. Works with or without a Prompt.
Error Messages	mapping	Error codes and descriptions that the model checks against the outcome.

Automatic starts with Google. It tries Exa only if Google fails before returning results and the server has an Exa key. Empty results do not trigger fallback. Selecting Google or Exa uses only that provider.

A Prompt receives the normalized results, including an empty list. Without a Data Schema, its answer is text. A Data Schema alone asks the model to return the results in that shape. The answer is stored in prompt_output. Invalid schemas are rejected when you save. Schemas with parameter references are checked after those references are filled in. A response that fails validation retries once before the block fails.

Leave Prompt and Data Schema blank to skip result processing. Error Messages still make an LLM call when configured on the block or workflow. Block entries override workflow entries with the same code. The model checks results, processed answers, and block failures against those descriptions, including failures before the search starts.

A successful search completes when no error code matches, including when it returns zero results. A matching code terminates a successful search. If the search or Prompt fails or times out, a matching code attaches to that failure without changing its status or system failure reason. If error detection fails, the block keeps the search outcome without a code.

The old no_results_error_code and no_match_error_code fields are deprecated. Saved values are read as Error Messages. They describe empty results and results that do not satisfy the Prompt. An explicit Error Messages entry with the same code takes precedence.

The output always includes:

• query: the rendered query, or the configured query if rendering fails.
• provider: google or exa.
• results: objects with title, link, snippet, display_link, and one-based position.
• total_count: the number of returned results.
• prompt_output: null, text, or the JSON value defined by Data Schema.
• raw_response.pages: provider responses, with credentials removed.

With Exa, a result’s snippet is empty when Exa has no cached copy of the page or when the request for page highlights fails.

For example, use {{ web_search_1_output.results }} in a downstream block. A later page failure keeps validated partial results and completes the search. A Prompt failure keeps those results but fails the block.

Failed and terminated outputs also include status, failure_reason, and errors. They carry a failure_category. For execution denials, such as insufficient credits, its value is null. Detected codes appear in the run’s errors list with their explanations. Enable Continue on Failure to let the next block inspect the output. The Search failure category applies to the run only when the block ends it.

A terminated Search block shows its failure reason in the timeline instead of the result count.

block_type: web_search
label: web_search_1
query: "site:example.com annual report"
provider: auto
num_results: 10
prompt: null
json_schema: null
error_code_mapping: null

Skyvern Cloud supplies provider credentials. Self-hosted installations configure SERPAPI_API_KEY and EXA_API_KEY in the server environment. Google pagination can use multiple API searches for a single block run.

---

Text Prompt

Send text to the LLM for processing without browser interaction. Useful for summarizing, transforming, or analyzing data between browser steps.

Field	Type	Description
Prompt	text	The text prompt to send to the LLM
Model	selector	Which LLM model to use
Data Schema	JSON	Expected output structure. Enable the checkbox to reveal the editor. Click Generate with AI to auto-suggest a schema.

File Parser

Parse PDFs, CSVs, Excel files, images, DOCX files, and ZIP archives.

ZIP inputs are unzipped, and the block outputs the extracted file list as file_name, file_path, and file_size. Data Schema is ignored for ZIP inputs; loop over the file list and pass each file_path to another File Parser block to parse individual files.

Very large files can exceed the parser’s time budget and fail the block. When a File Parser sits inside a loop, set On block failure (Advanced Settings) to Skip to next iteration so files that cannot be parsed are skipped and the rest of the list still runs.

Field	Type	Description
File URL	text	URL or parameter reference to the file
Data Schema	JSON	Schema for the extracted output. Enable the checkbox to reveal the editor.
Model	selector	Which LLM model to use

---

Control Flow

Loop

Repeat a sequence of blocks for each item in a list. Child blocks are placed inside the loop on the canvas.

Field	Type	Description
Loop Value	text	The list to iterate over. Use a parameter reference (e.g., {{ my_list }}) or a natural language description (e.g., “Extract links of the top 5 posts”) which triggers automatic extraction.
Continue if Empty	boolean	Mark the loop as complete if the list is empty

Inside loop blocks, use these reserved variables:

• {{ current_value }}: the current item
• {{ current_index }}: the iteration number (0-based)

Use Loop when the item list is already known before the loop starts, such as rows extracted from a table, files uploaded by a user, or URLs passed as agent parameters.

While Loop

Repeat a sequence of blocks while a condition remains true. Child blocks are placed inside the loop on the canvas.

Field	Type	Description
Condition	expression	Jinja2 template or natural-language prompt evaluated before each iteration

Inside while-loop blocks, use these reserved variables:

• {{ current_index }}: the iteration number (0-based)

Use While Loop for flows where the number of iterations is discovered during the run, such as pagination, polling, or retrying until a recoverable page state clears. For pagination, a common pattern is to extract a has_next_page boolean inside the loop, click Next, and let the next condition check decide whether to continue.

Conditional

Branch the agent based on conditions. The UI shows branches as tabs (A, B, C, etc.). Each branch has an expression that determines when it executes. Expressions can be Jinja2 templates or natural language prompts.

Field	Type	Description
Branches	tabs	One or more conditions. Each branch has an Expression field. Click + to add branches.
Else branch	auto	Automatically added. Executes when no other condition matches.

Example Jinja2 expression:

{{ extraction_output.price > 100 }}

AI Validation

Assert conditions using AI and halt the agent on failure. Useful for checking that a previous block produced expected results before continuing.

Field	Type	Description
Complete if…	text	Condition that means validation passed
Terminate if…	text	Condition that means validation failed

Additional fields in Advanced Settings:

Field	Type	Description
Error Code Mapping	JSON	Custom error codes for failure reasons
Input Parameters	select	Parameters to evaluate

Code

Execute custom Python code. Input parameters are available as global variables. Top-level variables in your code become the block’s output.

Field	Type	Description
Code	code editor	Python code to execute
Input Parameters	select	Parameters available as globals in the code

Password credential input parameters expose plaintext fields such as login_credential.username and login_credential.password. These values are readable by Python code inside the runner; they are not opaque handles. Use them only where needed, such as passing them directly to browser operations.

For one-time codes, use await login_credential.otp(), the supported pattern across both execution paths. Do not rely on .totp or the top-level otp(...) helper, whose behavior differs between paths. If no OTP source is configured, login_credential.otp() fails with a clear one-time-code-unavailable error.

await page.fill("#username", login_credential.username)
await page.fill("#password", login_credential.password)
await page.fill("#otp", await login_credential.otp())

The sandbox does not reject code that prints, returns, or transforms credential values. Before string values in a successful block output or a surfaced failure reason are persisted, Skyvern masks exact registered-secret matches; registered secrets of at least five characters are also masked inside larger strings. Derived or transformed values may no longer match. Do not expose credentials through output or errors. Opaque credential containment is tracked separately in SKY-11771.

Wait

Pause agent execution for a specified duration.

Field	Type	Description
Wait in Seconds	number	Seconds to wait (0–300)

---

Files

File Download

Navigate the browser to download a file.

Field	Type	Description
URL	text	Starting URL (optional)
Download Goal	text	Instructions for finding and downloading the file
Download Timeout (sec)	number	Seconds to wait for the download to complete

Additional fields in Advanced Settings:

Field	Type	Description
File Name	text	Custom filename for the download. Replaces the site’s filename, keeping the extension. Reference {{ original_filename }} to keep the site’s name instead — {{ fund }}_{{ original_filename }} prefixes it, which separates documents that several pages name identically.

Plus common browser fields.

Cloud Storage Upload

Upload downloaded files to S3 or Azure Blob Storage.

Field	Type	Description
Storage Type	select	Amazon S3 or Azure Blob Storage
Folder Path	text	Upload path/prefix (optional)

S3 fields

Field	Description
S3 Bucket	Bucket name
AWS Access Key ID	AWS credential
AWS Secret Access Key	AWS credential
Region Name	AWS region

Azure fields

Field	Description
Storage Account Name	Storage account name
Storage Account Key	Storage account key
Blob Container Name	Container name

---

Communication

Send Email

Send an email notification, optionally with file attachments from previous blocks.

Field	Type	Description
Recipients	text	Comma-separated email addresses
Subject	text	Email subject line. Supports parameter references — add {{workflow_run_id}} to tag the subject with the run, e.g. Your Run is Finished {{workflow_run_id}}.
Body	text	Email body. Supports parameter references.
Body format	select	Text (default) sends the body exactly as written. HTML renders the body as a formatted email and adds a plain-text version for clients that cannot show HTML; <script>, <iframe>, <object>, <embed>, <svg>, <math>, meta refresh directives, event-handler attributes and links whose scheme is not http, https, mailto, tel or cid (inline data:image sources are kept) are removed before sending. Mail clients strip inline SVG anyway, so use a raster image instead. {{ }} and {% %} are parameter syntax, so wrap literal braces in {% raw %}…{% endraw %}.
File Attachments	text	Path to files to attach

HTTP Request

Make an API call to an external service. Click Import cURL in the block header to populate fields from a cURL command.

Field	Type	Description
Method	select	GET, POST, PUT, PATCH, DELETE, HEAD, OPTIONS
URL	text	Request URL. Supports parameter references.
Headers	JSON	HTTP headers. Use Quick Headers to add common headers.
Body	JSON	Request body (POST/PUT/PATCH only)
Files	JSON	File uploads (POST/PUT/PATCH only)

Additional fields in Advanced Settings:

Field	Type	Description
Timeout	number	Request timeout in seconds (1–300, default: 30)
Follow Redirects	boolean	Automatically follow HTTP redirects (default: true)
Continue on Failure	boolean	Continue agent if the request fails
Save Response as File	boolean	Save the response body as a downloadable file
Download Filename	text	Filename for the saved response (shown when Save Response as File is enabled)

Human Interaction

Pause the agent and request human input. Optionally sends an email notification to reviewers.

Field	Type	Description
Instructions For Human	text	Instructions displayed to the reviewer
Timeout (minutes)	number	Minutes to wait before auto-continuing (default: 120 minutes / 2 hours)

Email notification fields:

Field	Type	Description
Recipients	text	Email addresses to notify (comma-separated)
Subject	text	Notification email subject
Body	text	Notification email body
Body format	select	Text (default) or HTML, as in the Send Email block. In HTML mode the run link, the instructions and the browser-session link are appended as an HTML footer inside </body> when the body is a full document.

Additional fields in Advanced Settings:

Field	Type	Description
Positive Button Label	text	Label for the approval action (default: “Approve”)
Negative Button Label	text	Label for the rejection action (default: “Reject”)