Automate Browser Tasks with AI-Guided Navigation

Chooses the cheapest correct Skyvern CLI command for a browser-automation task by classifying it first, based on a fixed decision table: a yes/no question uses skyvern browser validate (1 LLM call, the cheapest AI option); "what does the page show" uses skyvern browser extract (a dedicated extraction LLM with schema validation and caching); a single action on a known target uses deterministic skyvern browser click/type/select (0 LLM calls, fastest); a single action on an unknown target uses skyvern browser act (2-3 LLM calls, no screenshots, an economy accessibility tree); a same-page multi-step flow uses act when labels are clear or a chain of primitives when tighter control is needed; a throwaway exploratory trial uses skyvern browser run-task; and a multi-page or reusable automation is built as a skyvern workflow create/run, with one visually-reasoned, verified block per step and reusable run history. Decision rules layered on top of the table: use primitives instead of act whenever a selector, id, or XPath is already known; use validate instead of extract/act for boolean questions; use run-task for a genuine one-off; use workflow create for anything meant to be repeatable or scheduled; and never type a password directly - always authenticate through stored credentials via skyvern browser login.

Use it for AI-assisted browser navigation, extraction, form filling, login flows, or turning a one-off browser task into a reusable, verified workflow, especially when deterministic selectors aren't available and Skyvern's visual or accessibility-tree reasoning is needed to find page controls. Do not use it to bypass site access controls, rate limits, consent gates, or automation-prohibiting terms; confirm user intent before any page-changing action such as submitting, purchasing, deleting, or sending messages; and prefer deterministic selectors over AI-guessed ones for stable production flows, since AI actions can misread unlabeled or visually ambiguous controls. run-task is explicitly for exploration only, not recurring or multi-page production automation, which belongs in a workflow.

Every browser command needs a session first:

skyvern browser session create --timeout 30

Execution then follows the classification: skyvern browser validate --prompt "..." returns a boolean; skyvern browser extract --prompt "..." --schema '{...}' returns schema-validated structured data; skyvern browser click --selector "#submit-btn", type, and select execute deterministically with intent, selector, or hybrid targeting; skyvern browser act --prompt "..." handles same-page steps with unlabeled or unknown targets; skyvern workflow create --definition @workflow.yaml plus skyvern workflow run --id wpid_123 --wait and skyvern workflow status --run-id wr_789 build and run a multi-block, reusable automation whose first run uses AI and later runs replay a cached script 10-100x faster. Every page-changing action should be verified afterward with skyvern browser screenshot, a validate boolean check, or skyvern browser evaluate --expression "..." for JS-level state. Credentials are never typed directly - they are registered once with:

skyvern credentials add --name "my-login" --type password --username "user@co.com"

and referenced by ID through skyvern browser login --credential-id cred_123, with support for password, credit_card, and secret types plus Bitwarden, 1Password, and Azure Vault providers.

A CLI wrapping Skyvern's AI browser-automation engine, with an equivalent MCP interface (preferring an observe + execute pair for same-page multi-step UI work that the CLI doesn't expose directly); supports cloud sessions by default, local sessions for localhost or self-hosted targets, or connecting to an existing browser over CDP. All commands accept --json for structured output, and a large reference library (prompt writing, engines, schemas, pagination, block types, credentials, sessions, common failures, and more) backs deeper troubleshooting.

Developers automating websites that lack stable selectors or need AI judgment to find controls, who want to pick the cheapest correct primitive for each step - deterministic clicks where possible, AI only where necessary - and graduate a proven one-off task into a verified, reusable workflow.