# ToolBench from Arcade: 41,000 MCP Servers and the Tool Quality Crisis
ToolBench—a public benchmark from startup Arcade—has indexed 41,921 MCP servers and analyzed 218,422 tools. The result: only 0.5% of tools achieved an A rating or higher. 76.6% of tools got an F—less than 50 out of 100 points. This signals the immaturity of the ecosystem, where the number of servers is growing faster than quality.
MCP (Model Context Protocol) is an open standard for integrating AI agents with external tools. By March 2026, the protocol's SDK has been downloaded 97 million times per month. Backing from OpenAI, Google, Microsoft, and handover to the Linux Foundation. But ToolBench reveals: out of 42,000 servers, only about a thousand tools are reliable for production.
ToolBench Evaluation Criteria
The benchmark evaluates servers across four dimensions with different weights. The methodology draws on 54 agent tool patterns developed by Arcade based on 8,000+ production tools for enterprise clients.
For local servers:
- Tool description quality (50% weight): accuracy, completeness, alignment with actual capabilities.
- MCP protocol compliance (20%): JSON schema validity, call handling.
- Maintainability (30%): GitHub stars, commit frequency, license availability.
For remote servers, security is evaluated instead of descriptions:
- OAuth 2.0 and PKCE.
- Authentication correctness.
Scores are converted to grades: A+ (90–100), A (80–89), ..., F (<50).
| Rating | Points | Share of Tools |
|--------|--------|----------------|
| A+–A | 80+ | 0.5% |
| B–C | 50–79 | 23.4% |
| F | <50 | 76.6% |
Barriers for AI Agents
Problems go beyond servers. Enterprise platforms resist agent access. Slack, Workday, and LinkedIn restrict APIs for third-party AI. Arcade CEO Alex Salazar noted: even existing MCP servers are useless due to strict provider limitations.
This creates a vicious cycle: the protocol is growing, but agents can't use tools in real-world scenarios. Mid- and senior-level developers should consider these metrics when selecting MCP servers for production.
Key takeaways:
- 76.6% of MCP tools are unsuitable (<50 points) per ToolBench.
- Local servers are scored on descriptions (50%), protocol (20%), and maintainability (30%).
- Remote servers are scored on OAuth/PKCE and authentication instead of descriptions.
- Enterprise APIs (Slack, LinkedIn) block agents even when servers are available.
- Out of 42K servers, ~1K tools are production-ready.
Implications for Developers
ToolBench provides quantifiable metrics for filtering servers. Senior AI agent developers can integrate its API for automated validation. The MCP ecosystem needs a focus on quality: better descriptions, standardized security, and improved maintainability.
Growth to 97 million SDK downloads underscores the urgency: the protocol is in the Linux Foundation, but without quality, agents will remain sandboxed. Recommendation—prioritize A/B-rated servers per ToolBench in production stacks.
— Editorial Team
No comments yet.