Automate Browser Tasks with AI-Guided Navigation

Skyvern Browser Automation is a CLI judgment procedure for AI-powered browser work: every command is a runnable skyvern <command> invocation, and the skill's job is picking the right one for the task before running anything. It classifies the task first against a fixed table: a yes/no question uses skyvern browser validate (cheapest, 1 LLM call plus screenshots, max 2 steps); a page-content question uses skyvern browser extract with a JSON schema; a known-selector action uses deterministic Playwright primitives (click/type/select) with zero LLM calls; an unknown-target single action uses skyvern browser act, which reasons over an economy accessibility tree with no screenshots; a same-page multi-step flow uses act or a primitive chain; a one-off exploratory trial uses skyvern browser run-task; and a multi-page or reusable automation uses skyvern workflow create plus run, building one visually-verified block per step with reusable run history - the first run uses AI reasoning, subsequent runs replay a cached script that is 10-100x faster, and --run-with agent forces AI mode back on for debugging a flaky flow. Every page-changing action should be followed by a verify step - a screenshot for a visual check, validate with a yes/no prompt, or evaluate reading page state like document.title - rather than assuming the action worked.

Use it for AI-assisted navigation, extraction, form filling, login flows, or building a repeatable website workflow, especially when deterministic selectors are not available and Skyvern's visual/accessibility reasoning has to identify page controls. The decision rules are explicit: a selector, ID, or XPath in the prompt means use a primitive, not act; a plain yes/no question means validate, not extract or act; a one-off "try this once" trial means run-task, not a workflow; and anything meant to be reused, scheduled, or explicitly "set up" means workflow create, not run-task. Passwords are never typed directly - skyvern browser login always uses stored credentials instead. Do not use it to bypass site access controls, rate limits, consent gates, or terms that prohibit automation, and always confirm user intent before an automated action submits a form, makes a purchase, deletes something, or sends a message - browser automation changes remote state, and AI-driven actions can misread unlabeled or visually ambiguous controls, so deterministic selectors are the safer default for stable production flows.

Input is a browser session (cloud, local for localhost/self-hosted targets, or a CDP connection to an existing browser), created once and auto-attached to subsequent commands, for example:

skyvern browser session create --timeout 30

Output depends on the command: validate returns a boolean, extract returns schema-validated JSON built from screenshots and a dedicated extraction LLM, act/primitives perform the action, run-task returns the outcome of a one-off autonomous run, and workflow run/status return a run ID and status for a reusable, multi-block automation. Common failure fixes: if an action clicks the wrong element, add more context to the prompt or switch to hybrid targeting; if extraction returns empty, wait for content to load, relax required schema fields, or check the row count first; if login passes but the next step fails, confirm the same session is reused and add a post-login validate check; and if a prompt is overloaded, split it into one intent per command.

The Skyvern CLI and, alternatively, the Skyvern MCP server, where observe plus execute is preferred over act for same-page multi-step UI work since the CLI does not expose that pair directly. Targeting supports three modes on primitives: intent (AI finds the element), selector (deterministic CSS/XPath), or hybrid (selector narrows, AI confirms). Credentials are typed as password, credit_card, or secret, and can be stored directly or through a bitwarden, 1password, or azure_vault provider rather than ever being typed into a form field.

Developers automating web tasks that mix reliable, selector-based steps with visually ambiguous ones - login flows, form fills, data extraction, or multi-page wizards - who want the cheapest correct method chosen per step instead of defaulting every action to a full AI reasoning call.