Burnmeter

One dashboard for every AI coding agent — see exactly where your tokens go.

A local-first dashboard that reads the session logs your AI coding agents already write to disk and turns them into a real-time view of token usage, cost, burn-rate, and — where the agent actually logs it (e.g. Codex) — how close you are to your plan's rate limit. No API keys. No cloud account. No phone-home. Your data never leaves your machine.

ℹ️ Screenshot uses synthetic demo data (tests/generate_fixture.py) — your real numbers never leave your machine.

---

Why this exists

Neither Anthropic nor OpenAI exposes a real-time, per-token usage API for individual Pro / Max / ChatGPT-subscription plans. The in-app usage page and the CLI's /usage command are point-in-time only — not historical, not programmatic. So if you're on a flat-fee subscription, your local session logs are the only accurate source of truth for what you're actually burning.

Burnmeter reads them — ~/.claude/projects/**/*.jsonl and ~/.codex/sessions/**/*.jsonl — and answers the one question that matters: "Am I about to hit the wall?"

Features

• Binding-constraint hero — the single most-urgent number, front and center. Codex logs real rate-limit usage, so its hero is your rate-limit % (5-hour + weekly) — red ring at 90%+ means "about to get throttled." Claude Code exposes no account-level limit (Anthropic doesn't write one to disk), so its hero is burn-rate $ + live local usage — honestly labelled, never a guessed limit.
• Real per-model pricing — Opus / Sonnet / Haiku and GPT-5.5 / 5.4 / Codex / mini, each at its actual rate. Output costs 5–8× input — the model mix is the cost story.
• Cache-efficiency panel — cache reads cost ~10% of fresh input. burnmeter shows your hit rate and the dollars cache saved you vs. an un-cached run (often the "I'm getting $50k of value for $200" number).
• Per-project attribution — which repo is eating your quota.
• Burn-rate trend, daily spend, per-turn activity, model split — the full "where did it go" picture, with sparklines and clean charts.
• Multi-device — sync a second machine's logs in (e.g. via Syncthing) and see per-device breakdowns.
• Terminal statusline — burnmeter statusline prints a one-liner for your shell prompt or Claude Code's status bar (verdict · model · $/h · block left · errors · cache · today). Same numbers as the dashboard, no server needed.
• Tiny footprint — the long-running server holds ~30 MB; heavy parsing runs in a short-lived worker so memory never balloons.

Install

Python 3.10+, zero third-party dependencies (Chart.js / GridStack load from a CDN in the browser). Recommended: pipx for an isolated install.

pipx install git+https://github.com/cihanatak/BurnMeter   # PyPI release coming → `pipx install burnmeter`
burnmeter serve            # starts the dashboard and opens it in your browser
burnmeter desktop          # optional: adds a double-click "Burnmeter" desktop shortcut

Or with pip, or straight from a source checkout:

pip install git+https://github.com/cihanatak/BurnMeter
burnmeter serve
# — or, from a source checkout —
git clone https://github.com/cihanatak/BurnMeter
cd BurnMeter
python3 -m burnmeter serve

Multi-source (Claude + Codex) and a synced second machine:

burnmeter serve --extra-projects-dir ~/.claude/projects-pc --codex-extra-dir ~/.codex/sessions-pc

Running on another machine

The whole app is in this repo — clone and run. Everything else is machine-local by design (not committed):

git clone https://github.com/cihanatak/BurnMeter
cd BurnMeter
python -m pip install -e .          # free core, zero third-party deps
# Pro cross-device sync (optional): adds `cryptography`
python -m pip install -e ".[sync]"
python -m burnmeter serve --port 9876

• It reads that machine's own logs — ~/.claude/projects and ~/.codex/sessions (via Path.home(), so it works on macOS, Linux, and Windows: C:\Users\<you>\.claude). No data is copied between machines.
• Auto-start is OS-specific and not in the repo. macOS uses a launchd LaunchAgent; on Windows use Task Scheduler or just run python -m burnmeter serve. (Caches under ~/.claude / ~/.codex regenerate on first run.)
• Pro sync to link two machines: install the [sync] extra on both, run a relay somewhere (burnmeter sync-relay), then on each machine burnmeter sync login --relay <url> --token <same-token> with the same passphrase, and burnmeter sync push. Each shows up in the other's Cihazlar card. The relay only ever stores ciphertext; the passphrase never leaves your machines.

Terminal statusline

A one-line live readout for your prompt or Claude Code's status bar — verdict dot, active model, $/h burn, block time left, 24h errors, cache hit, today's spend:

burnmeter statusline
# 🟠 Sonnet 4.6 $48/h · blok 1h34m · 12err/24h · cache %97 · gün $42

burnmeter statusline --source codex   # Codex instead of Claude
burnmeter statusline --json           # machine-readable, same numbers

It computes standalone (no running server needed) and shares its logic with the /api/statusline HTTP endpoint, so the terminal line and the dashboard never drift.

Wire it into Claude Code (~/.claude/settings.json) so it refreshes in your prompt:

{
  "statusLine": { "type": "command", "command": "burnmeter statusline" }
}

How it works

~/.claude/projects/**/*.jsonl  ─┐
~/.codex/sessions/**/*.jsonl   ─┤→ parser → normalized UsageRecord
                                │            ↓ (in a short-lived worker)
                                └──────────→ analytics → report JSON
                                                         ↓
                                       local web dashboard (127.0.0.1)

No network calls except loading Chart.js from a CDN (vendor it for fully air-gapped use). Costs are client-side estimates based on published pricing — for billing, see your provider's console.

Privacy / local-first

Everything stays on your disk. Burnmeter binds to 127.0.0.1 only. There is no telemetry, no account, no upload. The optional Pro sync feature (see PRICING.md) is end-to-end encrypted and entirely opt-in.

Supported agents

Agent	Source	Status
Claude Code	~/.claude/projects	✅
OpenAI Codex	~/.codex/sessions	✅
Cursor / Copilot / Gemini / others	—	🛣️ planned

Adding an agent is a single parser module — see CONTRIBUTING.md.

Pricing

The local dashboard is free and open source forever. A Pro tier adds cross-device sync, unlimited history, and pre-limit alerts; a Team tier adds cost aggregation & attribution. Details in PRICING.md.

License

AGPL-3.0-or-later for the open-source core, with a commercial dual-license available for organizations. Free for individuals, including commercial use.