context-cost-line-item-analyzer

To analyze agent token spend, you decompose session costs into context-loaded-before-first-prompt versus output-generated-during-session. This breakdown classifies spend into startup overhead, system prompts, memory, MCP tool schemas, and skill descriptions to identify the main context cost drivers.

The best way to reduce context window overhead in CLAUDE.md chains is to rank context cost drivers by estimated token budgets. By pinpointing startup overhead and system prompts within the chain, you can apply targeted reductions to trim unnecessary context tokens from ongoing sessions.

MCP server configuration impacts token costs through tool registration schemas loaded into the context window. Analyzing these configurations reveals how much startup overhead and memory usage they consume, allowing you to measure their specific contribution to overall context-driven spend.

Yes, you can measure startup overhead and system prompt token usage by classifying spend loaded before the first prompt. This process isolates initial context costs from output-generated tokens, producing a prioritized list of items with estimated token budgets for targeted optimization.

To optimize SKILL.md descriptions and lower context token usage, you analyze them to surface their individual token budgets. This generates actionable reduction recommendations, helping you trim the context tokens these descriptions consume within the agent's startup overhead.

Your agent session consumes tokens before output because of context-driven costs loaded into the environment. Decomposing the session reveals that startup overhead, memory usage, MCP tool schemas, and system prompts are the primary drivers requiring targeted reductions to optimize spend.