ContextViz

Claude Code is powerful, but its context window is a black box. You know tokens are being spent; you rarely know on what. ContextViz opens the box: point it at your local ~/.claude/projects directory and see the exact composition of every context window — system prompt, MCP instructions, CLAUDE.md memory, conversation history, tool results, thinking blocks, and images — reconciled to the per-turn token total.

The Redundancy Radar surfaces content you're carrying more than once and prices the waste. The File Time Machine reconstructs every version of every file a session touched, including intermediate states git never captured. The Subagent Explorer breaks out the hidden cost of spawned Agent and Task runs: model, peak context, dollars spent, and tool call count. Everything is computed locally — ContextViz reads your own ~/.claude data and never sends it anywhere.

Cache Economics turns prompt caching from a hopeful checkbox into a measurable line item: cost, dollars saved, hit rate, and a per-turn waterfall. Live-tail mode keeps the X-ray refreshing as an active transcript grows, so you can watch context composition change in real time while Claude is running.

The design philosophy is deliberate scope: broad agent dashboards browse sessions and total up costs across tools. ContextViz goes the other way — a microscope for Claude Code specifically. Because it focuses on one agent, it can reconstruct things a breadth-first tool cannot: the exact composition of each context window, the carry-cost of redundant content, and every intermediate file version. Single-agent scope is the point; it's the deepest view into Claude Code there is.