100x-token-meter

See where your coding agent's tokens go. 100x-token-meter reads the local logs of Claude Code, Codex CLI, Gemini CLI and Hermes Agent, including open models on OpenRouter or Ollama, and writes one small summary file, token-meter.json. Drop that file on the Token Meter page to see your numbers in charts, along with your biggest token tax.

npx 100x-token-meter

It runs offline and only reads files. The summary holds totals, tax scores and a cost split by source only: tool, MCP server and skill names, never prompts, code, file names or paths.

Setup guide

1. Check you have Node.js 18 or later

node -v

If that prints nothing or a version below 18, install the LTS version from nodejs.org.

2. Run it on the computer where you use your coding agent

npx 100x-token-meter

npx downloads it the first time. It reads every tool it finds. You'll see something like:

Read 12,467 requests from 292 sessions (last 30 days).
  Claude Code: 283 sessions, 10,961 requests, $1091.89
  Codex CLI: 9 sessions, 1,506 requests, $68.67
About $1160.56 at API prices. Biggest tax: noisy output, $100.08.
Wrote /Users/you/token-meter.json. Drop it on the Token Meter page.

3. Load the summary

Open the Token Meter page, go to Your sessions, and drop token-meter.json there. The file is read in your browser and is never uploaded.

Flags

| Flag | Default | What it does | | --- | --- | --- | | --tool NAME | all | claude, codex, gemini or hermes to read one tool only | | --days N | 30 | How many days back to read | | --out FILE | token-meter.json | Where to write the summary | | --help | | Show usage |

npx 100x-token-meter --tool codex              # Codex CLI only
npx 100x-token-meter --days 7                  # just the last week
npx 100x-token-meter --out ~/Desktop/tm.json   # write it somewhere else

Where the logs are read from, and how to point somewhere else:

| Tool | Folder | Override | | --- | --- | --- | | Claude Code | ~/.claude/projects, ~/.config/claude/projects | CLAUDE_CONFIG_DIR (comma-separated for more than one) | | Codex CLI | ~/.codex/sessions, ~/.codex/archived_sessions | CODEX_HOME | | Gemini CLI | ~/.gemini/tmp/<project>/chats | GEMINI_CLI_HOME | | Hermes Agent | ~/.hermes/state.db | HERMES_HOME |

Hermes keeps a SQLite database, which needs Node 22.13 or later to read. On older Node the other tools still run and Hermes is skipped with a note. Hermes stores totals per session and model, not each request, so its sessions add to the totals and the model table but have no tax scores or split by source; by_tool lists what each tool could measure.

What it measures

Totals come from each response's own usage numbers, counted once per request. Costs use the bundled prices.json, taken from Anthropic's, OpenAI's and Google's pricing pages; the date they were checked is in the file. Costs are what the same tokens would cost on the API, even if you're on a subscription.

Open models are priced at OpenRouter's rates. A model you run on your own machine (Ollama, LM Studio) bills nothing, so it is priced as the same model hosted on OpenRouter, marked local in by_model, and summed in local. Ollama tags are matched only where the family and size match (llama3.2:3b is Llama 3.2 3B); others, such as qwen3:8b, which OpenRouter doesn't host, show as not priced.

| Tax | Rule | | --- | --- | | Noisy output | Tool results over 5,000 tokens (characters ÷ 3.5), written once and re-read on every later request | | Screenshots | Image blocks, by each provider's rule: Claude counts 28 px squares, OpenAI 32 px squares × 1.2, Gemini a flat 1,120 per image | | Startup | The new tokens on each session's first request: system prompt, CLAUDE.md / AGENTS.md / GEMINI.md, tools, MCP servers | | Cache breaks | Over 20K new tokens after a model change, or after a pause longer than the cache lifetime | | One model | Short jobs on the top model (10 requests or fewer, none writing 300+ tokens): what they cost over the same provider's small model |

Where it went, by source

by_source splits the same total by what was in the model's context: the startup (system prompt, memory file, tool list), your prompts, the replies, each built-in tool, each MCP server, each skill, subagents, and the tool's own notes.

Each request's real cost is shared out over its context in proportion to estimated size (characters ÷ 3.5): cache reads over everything already stored, cache writes and input over what is new, output over the blocks the model wrote. A Claude Code subagent's whole log counts as Subagents. The split is an estimate; its total matches the real one.

Troubleshooting

| Message | Fix | | --- | --- | | No … requests in the last 30 days | Read further back with --days 90, run it on the machine where you use the tool, or set the folder override above | | requests on … aren't in prices.json | Those models have no price yet, so they count as requests but add no cost | | Skipped Hermes Agent | Its log is SQLite; update to Node 22.13 or later | | npx: command not found | Install Node.js 18 or later from nodejs.org |

License

MIT