toolbench-mcp

FastMCP 3.1 helper for Arcade ToolBench: curated links, rescoring steps, Glama vs ToolBench notes, and an optional explanation of Arcade’s own MCP product (their hosted integrations runtime — separate from grading).

This project is not affiliated with Arcade. It does not call ToolBench HTTP APIs; the optional Playwright scraper loads public pages like a browser. You are responsible for complying with Arcade’s terms and for using the scraper only in line with the intended use section below.

---

If you have never heard of it, you are not alone — it is easy to miss.

ToolBench is a public website run by Arcade at [toolbench.arcade.dev](https://toolbench.arcade.dev/). It is part of Arcade’s ecosystem around MCP servers (the protocol many AI coding tools use to expose tools to models).

Roughly, ToolBench is where Arcade:

• Indexes and grades MCP servers that appear in their ecosystem (report cards, scores, methodology).
• Publishes methodology (how scores are derived), improve guidance, and a submit flow so maintainers can request rescoring after changes.
• Surfaces links to related Arcade docs (e.g. agentic tool patterns).

It is not the same thing as:

• Glama (glama.ai) or other directories — those are mainly discovery; ToolBench tends to emphasize transparent methodology and actionable report pages for servers Arcade tracks.
• Arcade’s “MCP product” (hosted integrations like Gmail/Slack through their runtime) — that is a separate product for using Arcade’s tools with auth; it does not replace reading your ToolBench report or improving your own server repo.

This repo (toolbench-mcp) gives you a small MCP server plus a local webapp so agents and humans can open the same links, follow rescoring steps, and optionally run a bundled Playwright scraper to save offline copies of ass