BurnRate

Cost & quota dashboard for Claude Code and Codex CLI — inside VS Code.

See your spend burn down in real time. Forecast when your Codex weekly quota will run out. Catch token spikes — even on routed models like GLM, Qwen, Kimi, or DeepSeek that other trackers price as $0.

All local. No API keys. No telemetry. No outbound network.

Why BurnRate

Most usage trackers answer "what did I spend yesterday?" BurnRate answers three questions the others don't:

1. When will I run out?

Codex has a weekly quota. BurnRate projects exhaustion from your local 7-day pace and flags when you're trending over the limit — before OpenAI's throttle hits mid-PR.

2. Which turn blew up my budget?

Spike detection surfaces the top cost turns in the range. For priced models, the threshold is USD. For unpriced or routed models, it falls back to token volume — so you still see the spike even when cost is $0.

3. What's actually happening when I route through GLM / Qwen / Kimi?

If you use Claude Code with a third-party proxy (common in 中国 users' setups), ccusage and most dashboards report $0 forever. BurnRate shows the token volume, lets you add pricing with two lines of config, and makes the invisible traffic visible.

Is this different from "Claude Code and Codex Assist"?

Yes — different category, not a competitor. That extension is a session history browser (diffs, full-text search, resume conversations) with usage tracking as a supporting feature. BurnRate is a dedicated cost & quota dashboard — we don't browse sessions, they don't forecast quota. Install both if you want both.

What It Reads

• Claude Code: ~/.claude/projects/*/<session>.jsonl
• Codex CLI: ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl

On Windows:

• Claude Code: %USERPROFILE%\.claude\projects
• Codex CLI: %USERPROFILE%\.codex\sessions

Built-in pricing covers Claude Opus / Sonnet series and OpenAI GPT-5.x / GPT-4.1 / GPT-4o. Everything else — including routed domestic models — is one burnRate.customPricing entry away.

Quick Start

Install from the VS Code Marketplace, then run BurnRate: Show Dashboard from the Command Palette.

Building from source:

npm install
npm run compile

Press F5 to launch the Extension Development Host.

Entry Points

BurnRate gives you three ways to reach the dashboard without hunting through menus:

• Activity Bar. A flame + bars icon in the left activity bar opens a welcome view with a single "Open Dashboard" button. Useful if you want BurnRate a click away at all times.
• Status Bar. A compact 🔥 $12.34 item in the bottom-right shows today's implied spend at a glance. Hover for a Claude / Codex split, click to open the dashboard. When you have no priced spend yet but have routed-model volume, it falls back to a token count (🔥 1.2M). Hide it entirely with burnRate.showStatusBar: false. Adjust refresh cadence with burnRate.statusBarRefreshSeconds (default 300, minimum 30).
• Command Palette. BurnRate: Show Dashboard and BurnRate: Refresh Data remain the canonical entry points.

Dashboard Features

Spend & usage

• Provider switcher: All Tools / Claude Code / Codex
• Total spend, implied API spend, tokens, cache-reuse ratio
• Project rows with spend bars and per-tool share in All Tools
• Model breakdown with priced / unpriced labeling

Heatmaps

• Calendar heatmap for month and all-time
• Hourly heatmap for today
• Theme-aware colors (works in dark + light)

Forecasting & alerts

• Codex weekly quota card with exhaustion projection from 7-day local pace
• Spike turn list (priced threshold + unpriced token-based fallback)
• Collapsible unknown-pricing banner with per-model ignore
• All-time overview with sparse-month placeholders and year grouping

Languages

• English and 简体中文 (burnRate.language setting, auto by default)

Routed / Unpriced Models

If you route Claude Code through a proxy to models like GLM-4.5, Qwen-Max, Kimi-K2, or DeepSeek-V3, they show up in the dashboard the moment you use them. Token volume is always tracked. To get dollar estimates, add two lines:

{
  "burnRate.customPricing": {
    "glm-4.5":     { "input": 0.00, "output": 0.00, "cache_read": 0.00 },
    "qwen-max":    { "input": 0.00, "output": 0.00 },
    "kimi-k2":     { "input": 0.00, "output": 0.00 },
    "deepseek-v3": { "input": 0.00, "output": 0.00 }
  }
}

All prices are USD per 1M tokens. Fill in the numbers from each provider's official pricing page — rates change often and BurnRate deliberately ships no live price feed. Prefix matching is supported (gpt-4o matches dated variants).

Don't want to price them? Add them to burnRate.ignoredUnpricedModels and they'll stop triggering the warning banner. Token-based spike detection still works.

Settings

burnRate.customPricing

Add pricing for models not covered by the built-in table:

{
  "burnRate.customPricing": {
    "my-gpt-alias": {
      "input": 2.5,
      "output": 10,
      "cache_read": 1.25,
      "reasoning_output": 10
    }
  }
}

Notes:

• All prices are USD per 1M tokens.
• reasoning_output defaults to output.
• Legacy cache_write is accepted as cache_write_5m.
• If a log only contains legacy cache_creation_input_tokens, BurnRate prices it at the 5-minute cache-write rate.
• If cache_write_1h is omitted, it falls back to the 5-minute rate.
• Prefix matching is supported.

burnRate.spikeThresholdUsd

Turns above this per-turn USD threshold are listed as priced spikes. Unpriced models still surface in spikes by token volume.

burnRate.ignoredUnpricedModels

Models in this list are hidden from the warning banner. They still count as unpriced ($0) unless you also add pricing for them.

burnRate.language

auto (default) follows VS Code's display language. Force English with en or 简体中文 with zh-cn. Changes apply live — no restart.

Settings Migration

Legacy claudeCostTracker.* settings are still read for one migration window:

• claudeCostTracker.customPricing
• claudeCostTracker.spikeThresholdUsd
• claudeCostTracker.ignoredUnpricedModels

Move them to burnRate.* before v3.0.

Accuracy Notes

BurnRate is a local-log estimate layer, not an official billing export.

• Unpriced models are counted as $0 until configured.
• <synthetic> Claude turns are excluded from spend totals (count shown in UI).
• Codex spend is shown as implied API cost even on flat-rate plans — so you can compare plan value vs raw API pricing honestly.
• Codex quota and projection mix two sources:
  • usedPercent comes from OpenAI rate-limit snapshots in the local Codex log
  • window usage, 7-day avg, and projection are derived from local log events
• Projection can undercount if old sessions were pruned or usage happened on another machine.
• Claude Code does not expose subscription quota in its local logs — that's why only Codex gets a Quota card.

Dedup details: Claude logs are deduped per-file by message.id. Codex logs are deduped within a file by (file, timestamp), and token_count events are associated with the preceding turn_context to recover the model name.

Log Retention

Claude Code prunes session logs after 30 days by default. To keep longer history, edit ~/.claude/settings.json:

{ "cleanupPeriodDays": 100000 }

Without longer retention, All Time only means "all logs still on disk".

Cross-Platform

BurnRate runs on macOS, Linux, and Windows. The repo CI builds on all three.

On Windows:

• Claude project-dir decoding is intentionally conservative; BurnRate prefers the recorded cwd field over guessing from encoded folder names.
• Path rendering keeps the drive letter when truncating long paths.
• Empty states and dashboard copy refer to the platform's default local log directories rather than Unix-only example paths.

Scope

BurnRate is intentionally narrow:

• Cost & quota dashboard, not a session browser
• Small built-in pricing table + user overrides for everything else
• No remote pricing feed
• No telemetry, no outbound calls, no temp files, no DB

If you need diff viewing, session resume, or full-text search across conversation history — pair BurnRate with a dedicated history viewer.

License

MIT