Skyvern-AI/skyvern

Skyvern – AI‑powered browser automation

What it is – Skyvern is an open‑source platform that lets you automate any website using large language models (LLMs) and computer‑vision instead of hand‑written selectors. It builds on the Playwright browser‑automation library and adds a thin SDK that accepts natural‑language prompts (e.g., “click the green Submit button”) to locate elements, extract data, validate page state, or run multi‑step workflows.

How it works – Skyvern runs a swarm of agents:

1. Perception – a vision‑enabled LLM looks at the rendered page and maps visual elements to actions.
2. Planning – the agents decide what steps are needed to satisfy the user’s high‑level prompt.
3. Execution – the chosen actions are sent to Playwright, which drives the browser.

Because the agents rely on visual cues rather than static CSS/XPath selectors, the automation is tolerant to layout changes and can work on sites it has never seen before.

Key features

• AI‑augmented Playwright commands – any Playwright action (click, fill, select_option, upload_file) can receive a prompt argument that lets the LLM locate the element by description.
• High‑level page API – page.act, page.extract, page.validate, and page.prompt expose natural‑language operations directly.
• Workflow builder – a no‑code UI lets you chain tasks (browser actions, data extraction, loops, email, HTTP calls, custom code, etc.) into reusable workflows.
• Live streaming – you can watch the browser viewport in real time, which helps debugging and manual intervention.
• Authentication support – built‑in login helpers, 2FA (TOTP, email, SMS), and integrations with password managers (Bitwarden, custom services).
• Model Context Protocol (MCP) – you can plug in any LLM that implements MCP, giving flexibility over providers.
• Cloud & self‑hosted options – a managed Skyvern Cloud service or a fully local stack (Python package + optional Docker‑Compose with Postgres).
• Performance – claims state‑of‑the‑art results on the WebBench benchmark (≈64 % accuracy) and top scores on “WRITE” tasks such as form filling and file downloading.

Getting started

1. Python install – pip install "skyvern[all]" (or skyvern for the API only).
2. Quick‑start – run skyvern quickstart (SQLite DB by default) or skyvern quickstart --docker-compose for a containerised stack.
3. Use the SDK – after launching a browser, call the AI‑augmented methods:

   await page.click(prompt="Click the login button")
   data = await page.extract("Get product name and price")
   await page.agent.run_task("Complete checkout for John Snow")
   

4. No‑code UI – open http://localhost:8080 to build and run workflows visually.

Typical use cases

• Robotic Process Automation (RPA) for internal tools where UI changes break traditional scripts.
• Data‑scraping / extraction from sites that lack stable APIs.
• Automated testing that needs to survive UI redesigns.
• Personal assistants that can control a browser you already have logged into (via Chrome remote‑debugging or tunnelling).

Where to find more – Full docs at https://www.skyvern.com/docs, a technical report on the WebVoyager evaluation, and community chat on Discord.

---

All information above is taken directly from the repository’s README; no additional features have been inferred.