MCP Makes Browser Automation Callable
David Mlcoch, co-founder & CEO of Asteroid, on browser automation and the last mile problem of AI
MCP does not replace browser automation, it turns browser automation into something an LLM can call like a normal tool. The key dependency is whether a real API exists. In places like healthcare, insurance, and supply chain, many core systems still only expose a website, so Asteroid uses a hosted browser agent to log in, click through forms, and enter data, then wraps that capability in MCP so upstream voice or chat agents can use it in a structured way.
-
The practical stack is API first, browser second. When an API exists, it is faster, cheaper, and more reliable. Browser agents are the last mile layer for portals with no programmable interface, especially old systems of record where staff still copy information from calls or emails into web forms by hand.
-
MCP mainly changes packaging, not the underlying work. Instead of asking an LLM to directly drive a browser session, Asteroid can expose browser actions through an MCP server, which makes a legacy website look more like a callable software tool even though the real execution still happens in the browser.
-
This is also how the market is segmenting. Browserbase and Stagehand serve developers who want raw browser infrastructure and frameworks. UiPath sells broader enterprise automation with heavier setup. Asteroid is aiming at operations teams that need repeatable browser work run at scale without building custom code for every portal.
The direction of travel is a layered agent stack where MCP becomes the standard control surface, APIs handle modern software, and browser agents cover the long tail of legacy software that still runs real industries. As long as important workflows live behind human only web portals, browser automation remains essential infrastructure rather than a temporary workaround.