AI BriefingOpenRouter Analysis: MCP Consumes 42x More Context Tokens Than CLI for GitHub Tasks
·2026.11.11 09:00
Key point
Testing on Claude Sonnet 5.5 shows MCP with 46 tools uses 59,417 input tokens versus 1,421 for CLI, with mitigation strategies reducing costs significantly.
Details
- MCP with 46 default tools consumes 59,417 input tokens ($0.1274) for a GitHub task, compared to 1,421 tokens ($0.0075) for CLI and 3,315 tokens ($0.0118) for Skills on Claude Sonnet 5.5.
- Tool definitions account for significant overhead; 46 GitHub tools add 19,436 tokens on Claude Haiku 5.5, while limiting to 2 tools reduces total task cost to 23,228 tokens.
- OpenRouter's Tool Search feature (Beta) reduces initial token load by deferring tool definitions, cutting Sonnet 5.5 costs from 59,417 to 30,334 tokens, though it does not reduce tool result size.
- Recommended usage: CLI for known commands, Skills for consistent CLI usage, MCP with scoped tools and result truncation for authentication needs, and Tool Search for large multi-server libraries.
- Prompt caching reduces token price by 90% but does not reduce context window occupancy; users should verify costs in their specific environment using the provided mcp_cost.py script.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Vercel Labs releases agent-desktop CLI for AI agents to control desktop apps via accessibility trees