tokencast

Lets you estimate LLM costs before running agent workflows, then learns from actual usage to improve accuracy over time. Exposes five MCP tools: estimate_cost takes size, file count, and complexity to predict token spend across research, architecture, and code steps with optimistic/expected/pessimistic bands. After work completes, report_session feeds actual costs back into a calibration system that adjusts future estimates. Ships with a Claude Code plugin that auto-triggers estimation after planning agents and records session data at shutdown. Also works standalone in Cursor, VS Code, and Windsurf. Useful when you're running multi-step agentic workflows and want to avoid surprise API bills or need budget guardrails before kicking off expensive tasks.