THE DUDE’S FIELD GUIDE
Browser automation without a paid model API
The quick answer
For steps you can specify in advance, start with Playwright, Puppeteer or Selenium running a local browser. Their basic scripted workflows do not need an LLM. Browser Use introduces model-guided actions, so you must separately evaluate the selected model, browser service and data flow.
This comparison summarizes documented capabilities. We have not run a comparative benchmark; fit depends on your workload and environment.
At-a-glance comparison
| Project | Best fit | Costs and services | Why consider it | Watch out for |
|---|---|---|---|---|
| microsoft/playwright | Repeatable browser tests | No model API required for local tests | Assertions, browser coverage and traces | Changing pages still require maintenance |
| puppeteer/puppeteer | Focused JavaScript browser scripts | No model API required for a local script | Browser control and screenshots | Manage the surrounding checks and compatible browser |
| browser-use/browser-use | Tasks requiring model-guided navigation | Model and cloud choices may incur charges | Agent uses a real browser | Local browser does not mean local or free inference |
Write the expected outcome first
A task such as opening your own page, completing a known sequence and checking a heading can be implemented as a script. A task with ambiguous instructions and changing navigation may benefit from a model, but that is a separate design choice. Start with a clear success condition and a reversible action.
Avoid confusing an assistant with the runtime
You can use Codex or Claude Code to write a browser script without making every execution call a model API. Inspect the generated script and dependencies. Conversely, a process running on your laptop may still send browser context to a remote model. Check actual configuration before describing the workflow as local.
Test failure as well as success
Try a missing element, delayed response and changed text. Require a visible error when an expectation fails. Save a trace or screenshot with private data removed, then repeat the same test after changing a selector. These evaluation steps are recommendations, not tests performed by The Repo Dude.
Common questions
Do I need Docker?
Not for every browser workflow. Read the selected package and operating-system instructions; container deployment is a separate choice.
Is a local browser agent free?
Not automatically. Browser execution, model inference, proxies and hosting are separate components with potentially separate costs.
Sources and review
- microsoft/playwright documentation
- puppeteer/puppeteer documentation
- browser-use/browser-use documentation
Check current installation instructions, licenses and service terms before choosing a tool. This guide’s review date changes only after an editorial recheck.
How we research and label evidence · Explore Browser & scraping