I wanted any AI Agent to access the web and perform tasks without costing me anything, which would normally require a browser running somewhere, but I didn’t want to run a headless browser or keep my laptop open. So I built an interface for agents that exists purely as a serverless function. The agent gets two primitives: ‘perceive’ and ‘act’.

You POST a URL, get back a structured map, act on an element by ID. That's the whole interface. And it costs almost nothing, at approximately $0.00003 per page load. The LLM call is actually more expensive than the browsing itself.

Works with Claude, GPT, Gemini, or any model that supports HTTP tool calls. The MCP server gives Claude Desktop native tools in three lines of config. Or grab the system prompt from the playground and paste it into any AI directly or deploy your own, it’s open source. Live playground: https://hollow-tan-omega.vercel.app/mirror GitHub: https://github.com/Badgerion/hollow