A single one-shot task is fine for a demo. The moment you need pagination, retries, or output you can trust, you want workflows. Skyvern exposes them through a visual editor, YAML, and a REST endpoint, and you can reuse the same structure across different sites.

What is a Skyvern workflow block?

A block is one unit of work: navigate, extract, run code, loop, validate, or call a task. Blocks pass data forward as JSON, which means the output of an extraction block can feed a conditional and then a download step. Typical production workflows run four to eight blocks and finish in under two minutes.

How do you build a pagination loop?

Create an extraction block that grabs the table rows, then a loop block that clicks the next-page control until it disappears. Give the loop a max-iteration guard, usually 50-100, so a broken site does not spin forever and burn API credits. This is the single most common source of runaway cost with Skyvern.

Can you call Skyvern from your own code?

Yes. The REST API accepts a task request and returns a run ID you can poll:

curl -X POST "$SKYVERN/api/v1/tasks" \
  -H "x-api-key: $KEY" -H "Content-Type: application/json" \
  -d '{"url":"https://portal.example.com","navigation_goal":"..."}'

Pair that with a webhook so your pipeline does not have to poll. In practice, webhooks plus a small queue is far cheaper than long polling because you stop paying for idle workers.

How do you cut vision-model costs?

• Cache the first two steps. Login flows rarely change, so a cached action can skip a vision call.
• Restrict screenshots to a region when the form is always in the same panel.
• Use a smaller, cheaper vision model for straightforward pages and reserve the flagship model for hard ones.
• Run extraction in bulk: one pass over a page beats five targeted micro-tasks.

On a 40-row table my test workflow fell from about $0.60 to $0.11 per run after caching the login steps and trimming screenshots. Numbers like that decide whether a job is viable at all.

Where does Skyvern still fall short?

It is not a speed tool. If your target is a JSON API, use the JSON API. Skyvern also struggles with canvas-heavy pages and aggressive bot detection, where a managed scraping service will do better. Treat it as a middle ground: more resilient than scripts, slower and pricier than direct HTTP.

Key Takeaways

• Workflows chain blocks and survive multi-step, paginated sites.
• Always cap loop iterations to prevent runaway API spend.
• Webhooks beat polling for cost and latency.
• Self-host the core and pay only for model tokens.

If you have ever maintained a Selenium script that dies every time a vendor redesigns their portal, you already know the pain Skyvern targets. Instead of telling it “click #submit-btn-3”, you tell it “download last month’s invoice” and it works out the clicks. Setup takes about 15 minutes if you already have Docker running.

What is Skyvern and what does it replace?

It replaces selector-based automation tools. The agent takes a screenshot of the page, asks a vision model what to do next, then acts. That loop costs money on every step because each screenshot is an API call to a vision model, so budget accordingly. A self-hosted run is basically free apart from your own model costs; the managed cloud charges per task.

How do you install Skyvern with Docker?

The fast route is the compose file from the repository:

git clone https://github.com/Skyvern-AI/skyvern
cd skyvern
docker compose up -d

That starts the UI on port 8080 along with a Postgres database. You will need an LLM key (OpenAI, Anthropic, or a compatible endpoint) in the .env file, plus a second key for the vision step unless you route both to the same provider. First boot pulled roughly 3 GB of images on my test box.

How do you write your first Skyvern task?

Give it a plain-English goal and a starting URL. The syntax in a Python task file looks like this:

await skyvern.run_task(
    url="https://example.com/portal/login",
    prompt="Log in with the saved credentials and download the newest PDF statement",
)

Run it from the CLI with skyvern run or trigger it from the web UI. The run viewer shows each screenshot and the decision made, which is genuinely useful when a step fails on step seven of a twelve-step form.

What does Skyvern cost?

The software is free and self-hostable. Your real cost is model tokens: expect a few cents per task on a mid-tier vision model. The hosted version uses usage-based pricing, so heavy daily jobs are better self-hosted. There is no per-seat licence on the open-source core.

Who should not use Skyvern?

If you need millisecond-level speed or you are scraping a simple static page, plain HTTP requests or Playwright will be 10-50x cheaper. Vision calls add 1-3 seconds of latency per page. Skyvern earns its keep on messy, login-gated, constantly changing sites.

How does Skyvern compare with Selenium and Playwright?

The honest summary: use Skyvern for the jobs you dread maintaining, and use code-based tools for the jobs that will never change.

Key Takeaways

• Skyvern is open source and self-hostable with Docker Compose.
• It uses vision models to read pages, so CSS selectors are not needed.
• Plan for vision API costs and slower per-page execution.
• Best fit: portals, government forms, and invoice downloads on sites you do not control.