Show HN: Hollow – serverless web perception for AI agents
TL;DR
Hollow is an open-source tool that lets AI agents browse the web through a purely serverless architecture, eliminating the need for persistent headless browsers. The interface provides two simple primitives—perceive and act—where agents POST a URL and receive a structured map to interact with. At roughly $0.00003 per page load, the browsing cost is actually lower than the LLM call itself. Compatible with Claude, GPT, and Gemini, it also offers an MCP server for Claude Desktop integration.
Nauti's Take
Hollow is proof that AI agent infrastructure does not have to be expensive or complex. Two simple primitives—perceive and act—beat bloated browser automation frameworks.
Nauti sees this as a sign that the bottleneck for agentic web access is shifting from cost to creativity.
Context
Browser-based web access is one of the biggest cost and complexity drivers in autonomous AI agents – Puppeteer or Playwright require running infrastructure that costs money and maintenance. Hollow solves this through radical simplification: no persistent process, no VM, just a serverless function. The pricing model makes web perception practically free for hobby projects and small productions.
For anyone building agents that browse, research, or fill out forms autonomously, this is an extremely low-barrier entry point.
Sources