Mantis

About

Mantis is a browser interface for AI agents that returns web pages as short numbered text instead of screenshots. It runs as an MCP server for Claude Code, Codex, Cursor or custom agents, numbers each page control, and returns only what changed after each action. CAPTCHA, 2FA and payment steps are handed to the human user. The authors report that on long tasks it used 29-47% fewer input tokens and about a third less time than Playwright MCP's accessibility tree approach, with accuracy within one run.