Agent Cost Mcp

Reads your local Claude Code session JSONL files and surfaces tool-by-tool cost attribution, daily spend trends, and retry loop detection without sending anything to the cloud. Exposes eleven tools including get_tool_usage for per-tool burn analysis, get_subagent_tree for parent/child cost rollup, get_cost_forecast for forward-looking estimates, and configure_budget for hard caps with webhook alerts. Use it when Claude Code's built-in /cost command gives you a number but you need to know which specific tool, branch, or runaway pattern caused it. Runs entirely via npx with zero setup, parsing existing logs from ~/.claude/projects/ to answer forensic questions about where your tokens went and what should change next run.