ToolBench: Grading Every MCP Server from A+ to F
Most MCP servers are not good. That’s not a hot take - it’s measurable. ToolBench is Arcade’s public MCP quality benchmark: we index MCP servers across the ecosystem and grade each one from A+ to F, with a full dimension-level breakdown of why.
The scoring runs on two models:
- Local servers: definition quality (50%) - tool naming, descriptions, parameter schemas, composability - plus protocol compliance (20%) and supportability (30%).
- Remote servers: protocol compliance (40%), security (30%) - OAuth 2.0, PKCE, transport security - and supportability (30%).
The point isn’t to dunk on anyone. Every grade links to actionable fixes drawn from Arcade’s 54 Agentic Tool Patterns - the same lessons behind my tool abstraction talk. The most common failure across the ecosystem is also the most fixable: missing tool descriptions, which leave agents guessing and picking the wrong tool. Fix your server, submit it for rescoring, and watch the grade move. There’s a Scoring API too, built on the same analysis used to evaluate 35,000+ servers.
Nate and I dig into ToolBench and the MCP Debugger in The Dark Side of MCP Servers.
- Benchmark: toolbench.arcade.dev
- Methodology: How we score every MCP server
- Improve your tools: patterns and top ecosystem issues
For more about Arcade check out the website at https://www.arcade.dev/ and the GitHub https://github.com/arcadeai/