Tool use via Model Context Protocol (MCP) — 36 real MCP servers, 220 tools, 1,000 human-authored tasks requiring 3-6 tool calls. Tests whether models can identify and orchestrate the right tools without being told which to use.
Uses the emerging MCP standard; tests real tool discovery and orchestration; open-sourced evaluation framework.
MCP is relatively new; tool landscape may shift as the protocol evolves.
Higher is better. Percentage of tasks completed correctly using MCP tools.
MCP is relatively new.
| # | Model | Score |
|---|---|---|
| 1 | 90.7 | |
| 2 | 88.1 | |
| 3 | 87.2 | |
| 4 | 87.2 | |
| 5 | 85.8 | |
| 6 | 85.4 | |
| 7 | 85.4 | |
| 8 | 84.5 | |
| 9 | 84.2 | |
| 10 | 83.6 | |
| 11 | 83.3 | |
| 12 | 82.3 | |
| 13 | 82.2 | |
| 14 | 82.2 | |
| 15 | 79.1 | |
| 16 | 78.2 | |
| 17 | 76.8 | |
| 18 | 75.8 | |
| 19 | 75.3 | |
| 20 | 70.8 | |
| 21 | 70.6 | |
| 22 | 69.4 |