Token Cost Calculator

Know exactly how much you're spending on API calls — before the bill arrives.

You're deep into a Claude Code session, hammering out features for hours, and suddenly wonder: "Wait, how much did today cost me?" Instead of logging into your Anthropic/OpenAI/DeepSeek billing dashboard and digging through usage charts, you open this plugin and see the answer in one click: $2.37 today, $18.52 this month, $143.80 since day one.

Token Cost Calculator is a CloudCLI tab plugin that reads your real Claude Code session transcripts (JSONL files stored locally at ~/.claude/projects/) and turns raw token counts into dollar amounts. It covers everything — main conversation turns, sub-agent sessions, cached prompt hits, thinking tokens from reasoning models — and gives you per-day, per-project, and all-time breakdowns with zero manual work.

---

What it does (in plain English)

• Auto-detect your real usage. No copy-pasting from a billing page. The plugin scans the JSONL transcript files Claude Code already writes to your disk, extracts actual input_tokens, output_tokens, cache_read_input_tokens, and cache_creation_input_tokens, and multiplies by the price you set.
• Five ways to slice the data. Manual entry for quick estimates. Auto-detect for the current session. Project mode for "how much did this entire repo cost me?" All-time for total spending. Daily mode with a date picker (defaults to today) so you can check "what did I spend last Tuesday?"
• Built-in model prices. Claude Opus 4.7, Sonnet 4.6, Haiku 4.5, GPT-4o, GPT-4o-mini, Gemini 2.5 Pro/Flash, DeepSeek V4 Pro — pick from the dropdown or type your own custom pricing per 1M tokens.
• Handles the weird stuff. DeepSeek (and other reasoning models) emit multiple content blocks per turn (thinking + text + tool_use), each written as a separate JSONL entry with duplicate message.usage. The plugin deduplicates by message.id so you're not 3x overcounting. Sub-agent sessions (from Claude Code's Task tool) are scanned too.

---

Install — 3 steps, no config files

Step 1 — Clone into your plugins folder

cd ~/.claude-code-ui/plugins
git clone https://github.com/NightmareAway/cloudcli-plugin-token-cost-calculator.git token-cost-calculator

Step 2 — Install and build

cd token-cost-calculator
npm install
npm run build

Step 3 — Restart CloudCLI

Close and reopen CloudCLI (or reload the plugins panel). The Token Cost Calculator tab appears automatically.

That's it. No API keys, no .env files, no configuration. The plugin reads whatever Claude Code already stores on your machine.

---

The five modes

Mode	Best for	What you see
Manual	Quick estimates, hypothetical "what-if" scenarios	Enter token counts yourself, tap quick-add buttons (+1K / +10K / +100K / +1M / +10M)
Auto-detect	The session you're in right now	Reads the active JSONL transcript live, includes all turns and sub-agents
Project	"How much did this whole repo cost?"	Sums every session under one project folder, lists each session with model name
All-time	"What have I spent since installing Claude Code?"	Aggregates all projects with colored progress bars — one bar per project
Daily	"Show me just today" or "what about last Wednesday?"	Day-by-day navigator (◀ ▶ buttons), defaults to today, click any day in the list

All modes share the same Price per 1M Tokens panel and Cost Breakdown at the bottom. Switch modes without losing your price settings.

---

Keywords & technical details

This plugin touches a bunch of topics people search for. Here's what's under the hood:

• Claude Code plugin / CloudCLI plugin — Built on the CloudCLI plugin architecture (manifest.json + mount/unmount lifecycle hooks). Runs a Node.js HTTP subprocess for backend logic.
• LLM token cost tracking / API spend calculator — Multiplies input_tokens / output_tokens / cache_read_input_tokens by per-1M-token prices. Supports Anthropic, OpenAI, Google Gemini, and DeepSeek pricing models.
• Prompt caching cost — Separately tracks cache_creation_input_tokens and cache_read_input_tokens, billed at the lower cache rate. This is often the silent cost driver in long conversations.
• JSONL session transcript parsing — Reads ~/.claude/projects/<encoded-path>/<session-uuid>.jsonl files line by line via Node.js readline. Filters "type": "assistant" entries and extracts message.usage.
• Message deduplication — Reasoning models (DeepSeek V4 Pro, and potentially Opus with extended thinking) emit multiple assistant JSONL rows per API call — one per content block (thinking, text, tool_use). Each row carries identical cumulative usage. The plugin tracks message.id in a Set<string> and skips duplicates.
• Sub-agent / tool-use sessions — Claude Code's Task tool spawns sub-agents whose transcripts live under <session-dir>/subagents/*.jsonl. The plugin recursively scans and includes these.
• Path encoding quirks (Windows) — Claude Code encodes project paths like C:\Users\... → C--Users-... (replacing \, :, spaces, and underscores with -). The plugin normalizes and reverse-maps these paths.
• Daily token aggregation — Groups usage by timestamp field (ISO 8601, extracted to YYYY-MM-DD). Frontend day navigator lets you step through dates or jump to today.
• Persistent state via localStorage — Mode, custom prices, and manual token counts survive page reloads.

---

Project layout

token-cost-calculator/
├── manifest.json        # CloudCLI plugin metadata (slot: tab)
├── package.json
├── tsconfig.json
├── icon.svg
└── src/
    ├── index.ts         # Frontend: 5-mode UI, day navigator, dark/light theme
    ├── server.ts        # Backend: HTTP API, JSONL reader, token aggregation
    └── types.ts         # PluginContext / PluginAPI / PluginModule types

API endpoints (backend)

Endpoint	Returns
GET /presets	8 built-in model prices
GET /session-usage?projectPath=...&sessionId=...	One session's token counts + model
POST /calculate	Cost breakdown for given tokens & prices
GET /project-usage?projectPath=...	All sessions under one project
GET /all-usage	Every project, ranked by total tokens
GET /daily-usage	All sessions grouped by YYYY-MM-DD

Develop

npm install
npm run dev     # tsc --watch
npm run build   # tsc

MIT.