Resend Meter: see how much of a coding-agent bill is the same system prompt sent again

Resend Meter's core pieces are already free and open source: a localhost proxy in front of the Anthropic or OpenAI API that logs tokens, cache hits and request diffs. tokdiet, claude-tap, tokentap, claude-thermos and claude-code-cache-fix all do some version of this, and ccusage-style log readers cover basic cost tracking with no proxy at all. tokdiet even pairs metering with context compaction and a shadow eval, which is roughly Resend Meter plus Prompt Diet in one small project. None of the results showed any tool doing the exact framing here: assigning every step's input tokens to system prompt, tool schemas, history and new content, and showing what share of the bill is re-sent bytes at cache-read versus full price as a per-session waterfall. That could still stand out as a sharper, better-designed view, but it is a feature that claude-tap or tokdiet could add quickly, and Anthropic's own /cost command is another place it could land. Prompt Diet looks a little less covered. There are CLAUDE.md trimmers that rely on rough rules, and there are generic eval harnesses, but I found nothing that cuts CLAUDE.md, AGENTS.md and the MCP tool list and then automatically replays saved tasks to prove the pass rate holds. tokdiet's shadow eval is the one overlap to look at closely. The most defensible way in is probably one cross-agent tool for Claude Code, Codex and Cursor-style agents that connects attribution to proven cuts, with a hard dollar-savings report. Even then it would be hard to charge individual developers in a space this full of free tools, so small teams, or a polished free tool feeding a paid team dashboard, look like the more realistic buyers.

• tokdiet

  The closest match to both ideas combined. It is a local reverse proxy that sits between Claude Code (and other agents) and the model APIs. It meters tokens and USD cost, compacts context to cut pay-per-token spend, and runs a shadow eval to show quality held. It also has a Claude Code plugin with a metering hook (https://www.claudepluginhub.com/plugins/agiwhitelist-tokdiet). It is small, about 68 stars, with the last push in June 2026.

• claude-tap

  A local proxy and trace viewer that works with Claude Code and many other agent CLIs. It records system prompts, conversation history, tool calls, token usage and request-to-request diffs, which already shows a lot of the repeated content between steps.

• claude-thermos

  A loopback reverse proxy for Anthropic traffic. It groups requests into sessions and cache-prefix lineages by model, tool set and system text, and it keeps the cache warm by replaying the last request.

• claude-code-cache-fix

  A local proxy or Node module that reports cache hit rates and telemetry for Claude Code. It is focused on raising cache reuse.

• tokentap

  A local proxy and dashboard that covers Claude Code, Codex and Gemini CLI.

• ccusage

  A widely used CLI that reads Claude Code's JSONL logs and reports cost by day, month and live session block. It does not split each step into prompt, tools and history.

• ccwhy

  A Rust tool that explains where Claude Code tokens went and suggests ways to reduce them. It overlaps with both the attribution idea and the suggestion idea.

• claude-trace-replay

  Turns Claude Code JSONL logs into a replayable view of agent flow and token usage.

• optimize-claude-md (mthines)

  A skill that audits CLAUDE.md files and ranks suggested trims. It has audit, interactive-trim and extract modes. It does not replay tasks to check that quality holds.

• cco-claudemd (claude-context-optimizer)

  Analyzes CLAUDE.md for token bloat, duplicate lines, filler wording and long code blocks, and suggests a fix for each.

• agent-eval (everything-claude-code)

  Runs coding agents on tasks defined in YAML, each in its own isolated git worktree, and reports pass rate, cost and consistency. It could serve as the replay harness for Prompt Diet, but it has no built-in way to A/B a CLAUDE.md change.

Checked 2026-10-25