Browser automation has changed. Instead of brittle CSS selectors that break every redesign, today's tools pair a real browser with a language model that reads the page and decides what to click. The good news: the best of these frameworks are open source with genuinely usable free tiers. Here are the five worth your time in 2026, and how to run each one for zero dollars.
The landscape in 30 seconds
AI browser tools split into two camps. DOM-driven tools (browser-use, Stagehand, Playwright MCP) read the page's structure and accessibility tree — fast, deterministic, cheap. Vision-driven tools (Skyvern) look at screenshots like a human — slower and pricier, but resilient to sites that fight selectors. Pick your camp by the target: stable web apps favor DOM, hostile or constantly redesigned sites favor vision.
1. browser-use — the open-source benchmark leader
browser-use is the most-starred project in this space (over 80k GitHub stars) and the standout on benchmarks, scoring about 89% on the WebVoyager task suite. It is a Python library under the MIT license: describe a goal in plain language and an LLM drives a real browser through the DOM and accessibility tree.
The library itself is free. Self-hosted runs cost you only the LLM tokens — and you can cut even that to zero by pointing it at a local model like Qwen or Llama through Ollama. A basic agent looks like this:
pip install browser-useimport asyncio
from browser_use import Agent, ChatOpenAI
async def main():
agent = Agent(
task="Find the cheapest direct flight from Lahore to Dubai next month and list the price",
llm=ChatOpenAI(model="gpt-4o-mini"),
)
result = await agent.run()
print(result)
asyncio.run(main())Best for: exploratory or one-off tasks, fast prototypes, general web navigation. Note that per-step LLM calls add up — budget roughly $0.02–$0.30 per task with a hosted model, or swap in a local model to make it truly free.
2. Stagehand — the TypeScript framework for repeatable flows
Stagehand, from the Browserbase team, is the TypeScript answer to browser-use. It sits on top of Playwright and lets you write natural-language commands (act("click the login button")) that fall back to explicit Playwright selectors wherever you need determinism. That hybrid design is its superpower: the AI handles the fuzzy parts, and your code stays debuggable.
Two practical advantages: it caches successful runs so repeated workflows replay with zero LLM calls, and it supports the Model Context Protocol (MCP), which means AI coding assistants can drive it directly. The framework is free and open source (MIT); you pay only for the model behind it and whatever browser infrastructure you run.
npm install @browserbasehq/stagehandBest for: production flows that must repeat reliably — onboarding scripts, recurring data pulls, self-healing test suites. Token costs per task typically land between $0.01 and $0.10, and caching pushes repeats toward zero.
3. Skyvern — vision-first for sites that fight back
Skyvern reasons over the rendered page visually instead of the DOM, so it keeps working when layouts change or when selectors are deliberately hostile. It defines workflows as goals, fills forms, handles CAPTCHAs, and even manages 2FA in its cloud tier. The core is open source and self-hostable via Docker, and the cloud tier includes a free credit allowance to get started.
The trade-off is cost: vision tokens are heavy (tens of thousands of tokens per workflow), and each step is slower than DOM-based tools. Use Skyvern when layout resilience matters more than speed — legacy portals, frequently redesigned dashboards, form-heavy internal tools.
Best for: volatile UIs, form-heavy workflows, and teams that need no-code workflow definitions alongside the API.
4. Nanobrowser — the free Chrome extension
Nanobrowser takes a different route: it is an open-source Chrome extension (13k+ stars) that brings an AI agent into the browser you already use. It is free — you bring your own LLM key, and that key can be a local model via Ollama, which makes the whole setup cost nothing. No servers, no Docker, no cloud accounts: install the extension, configure a model, and give it tasks in the sidebar.
Best for: personal productivity, one-off research tasks, and anyone who wants AI browser automation without touching infrastructure.
5. Playwright MCP — the fully free deterministic option
Not every task needs an LLM deciding steps at runtime. Playwright MCP exposes a real browser through the Model Context Protocol using the accessibility tree, letting AI assistants like Claude Code click, fill, and verify directly. There is no per-step model billing inside the tool itself, it is entirely free, and it shines in CI/CD pipelines and AI-assisted development workflows where determinism beats autonomy.
Best for: testing, CI/CD, and developers who want an AI assistant to drive a browser with surgical precision rather than free-roam autonomy.
How to actually pay $0
"Free tier" means different things, so here is the honest map:
- Self-host + local LLM: browser-use, Stagehand, and Nanobrowser all run against Ollama models. The software is free, the model runs on your hardware, and the total marginal cost is electricity.
- Cloud free tiers: Browserbase offers free browser hours, Steel gives 100 free hours of cloud browser infrastructure, and Skyvern's cloud includes a free credit allowance — enough to prototype before committing.
- Free framework, paid model: Stagehand's MIT framework costs nothing; you only pay the LLM provider per token, which drops toward zero on cached repeat runs.
If your target sites use aggressive bot detection, budget for proxies or a managed cloud tier eventually — the free core deliberately omits that layer.
Which one should you pick?
- Python team, one-off or exploratory tasks → browser-use
- TypeScript team, repeatable production flows → Stagehand
- Hostile or constantly changing sites → Skyvern
- Zero infrastructure, personal use → Nanobrowser
- Testing and CI, determinism first → Playwright MCP
Start with the free tier or a self-hosted setup, measure cost per task on your real workload, and only then decide whether a paid cloud tier earns its keep.