We analyzed 1,945 tools across 238 servers in the MCP Registry to understand how tools are distributed.

MCP Tool Distribution Brief

We analyzed 1,945 tools across 238 servers in the MCP Registry to understand how tools are distributed.

Distribution Snapshot

• 1 tool → 41 servers (17.2%)
• 2–5 tools → 113 servers (47.5%)
• 6–10 tools → 35 servers (14.7%)
• 11–20 tools → 26 servers (10.9%)
• 21+ tools → 23 servers (9.7%)

Key Insights

• Nearly half of all servers are small bundles (2–5 tools), suggesting that specific, targeted use cases are driving adoption.
• A meaningful long tail exists: 23 servers host 21+ tools, effectively becoming “tool platforms.”
• The 17% of servers with just one tool highlight highly specialized, focused utilities.
• The distribution raises an AI-UX question: do AIs perform better with larger tool sets, or with smaller, curated collections optimized for reliability?

Context from Recent Studies

• MCP-Universe Benchmark (Aug 2025): As tool count and interaction depth rose, every major LLM struggled more. Top models couldn’t break 44% accuracy, showing that complexity degrades performance.
• Microsoft Research – Tool-Space Interference: Found that performance drops when servers expose too many tools, especially with heavy outputs or overlapping definitions. This supports the “limited/curated sets” hypothesis with hard data.
• Security Research: Larger tool sets widen attack surfaces (tool poisoning, rug pulls). Limiting tools is not just about performance—it’s also basic security hygiene.

Takeaway
Current evidence suggests that tight, specific use cases lead to better performance and security, while overly broad tool sets can reduce effectiveness. The optimal balance between breadth and precision remains an open frontier for MCP design.