Perzian has released Mantis, a Model Context Protocol (MCP) server designed to let AI agents browse the web using short, numbered text representations instead of screenshots or accessibility trees. The tool, currently in v0.1.x beta, claims to reduce end-to-end wall time by about a third compared to the standard Playwright MCP implementation, while simultaneously reducing input token usage by 29-47% on long tasks.

Text-First Browsing Architecture

Unlike traditional browser automation tools that rely on visual screenshots or the DOM's accessibility tree, Mantis converts web pages into a compact, numbered text format that LLMs can parse more efficiently. The project supports integration with major coding agents like Claude Code, Codex, and Cursor, requiring only two npx commands for installation. By eliminating the heavy token cost of image data and verbose HTML structures, Mantis aims to lower the barrier for agents performing complex, multi-step web navigation.

Benchmark Results and Caveats

Internal benchmarks conducted on public sites using gpt-6-luna and claude-haiku-5-5 show significant efficiency gains. On long tasks, Mantis reduced median input tokens from 345k to 244k for Luna and cut median time from 66s to 42s compared to the accessibility tree baseline. However, the project discloses several limitations: it cannot handle canvas-based inputs (like WebGL or maps), requires human intervention for login walls, and was tested primarily on read, navigate, and form-fill tasks. The comparison against screenshot baselines showed mixed results, with screenshots using fewer tokens for Haiku.

Key Takeaways

  • Mantis reduces median input tokens by 29-47% on long tasks compared to Playwright MCP's accessibility tree approach.
  • The tool is currently in v0.1.x beta and lacks support for canvas, WebGL, and drawing applications.
  • Installation is streamlined for Claude Code users via a simple npx command.
  • Benchmarks show about a third less time for long tasks, though short-task time comparisons are noted as indicative due to build differences.

The Bottom Line

Mantis offers a compelling efficiency win for token-heavy agentic workflows, but its inability to handle canvas and visual inputs means it is currently a specialized tool for text-heavy sites rather than a universal browser replacement.