Browserbase provides headless browsers as a hosted service, and its official MCP server lets an agent drive them. The pitch is straightforward: browser automation is useful but running browsers is a nuisance, especially several at once, so let someone else run them and connect over the network instead.
What it actually does
The server creates and controls remote browser sessions. Claude can navigate, interact with pages, extract content and take screenshots, all against a browser running in Browserbase’s infrastructure rather than on your laptop. It supports Playwright underneath and Stagehand on top, which lets the agent describe an action in plain language instead of committing to a selector. Sessions are recorded, so a run that misbehaved can be watched back.
Practical patterns:
- ‘Scrape these forty product pages and give me a table of prices.’
- ‘Log into the staging site and walk through the signup flow, then show me the replay.’
- ‘Run this check against the live site every morning and tell me if the layout moved.‘
Why use it
Local headless browsers are fine for one task on one machine. They stop being fine when you want parallelism, when the job should run whether or not your laptop is open, or when you need to explain to someone why an automated run did something odd. Session replay in particular turns agent browser work from a black box into something you can debug.
Gotchas
Sessions cost money and agents are not naturally frugal; an unbounded loop over a large list will run up a bill. Set limits on both sides. There is also a latency floor from the network hop, so this is not the fastest option for a quick single-page fetch, where the local Fetch server would do. And you are sending page content through a third party, which matters if the pages are sensitive.
