Every token you burn. Every machine. One phone.
Track your AI coding agents across Windows, WSL and Linux:
- Spend: dollars and tokens per agent, model, workspace and session
- Cache hits: how much of your input came from cache
- Limits: vendor quotas (Claude, Codex, Z.ai and more) with reset countdowns
- Health: RAM, GPU and temperatures per machine
No servers. No accounts. No telemetry.
Light or dark, your call. Demo data: machines, usage and prices are made up.
machine: tokscale + burn-events → burn-report daemon ← Tailscale → phone: Expo + SQLite
- Machines are the backend: each runs
burn-report daemon, a read-only server (plus one opt-in control, below). - tokscale parses and prices ~50 AI clients.
- The phone pulls over Tailscale and keeps an offline copy.
- Machine offline? Its history is still there.
1. Join one tailnet. Install Tailscale on every machine and your phone.
2. Run one line on each machine.
Linux / macOS / WSL:
curl -fsSL https://github.com/ensevengg/burn/releases/latest/download/install.sh | shWindows (PowerShell):
irm https://github.com/ensevengg/burn/releases/latest/download/install.ps1 | iexIt installs burn-report, starts it now and with the machine, and prints a QR code. On Windows it offers to set up each WSL distro too, including the mirrored networking WSL needs to be reachable. Run it again to upgrade.
3. On the phone. Install burn.apk, tap Connect my machines, then Machines → + → Scan QR code. Your other machines appear in the same sheet: tap each one.
Lost the QR code? burn-report pair shows it again (add --port if you started the daemon with one). Just looking? Tap Explore with demo data.
Needs Bun and a Rust toolchain.
bun install
cargo install --path crates/burn-events # once per machine
bun run reporter -- install-service # or `daemon` to run in this terminal only- Rename a machine:
burn-report init --slug my-machine --name "My Machine". Every flag is optional. - One slug per install: Windows, WSL and each dual-boot side. By default Windows gets a
-windowsslug and each WSL distro a-wsl-<distro>slug. - Same box? Share a
--host-group. Windows and WSL share their hostname automatically. - Ports: 8787, or the first free port from 8788 to 8790 for a fresh WSL install (mirrored networking gives every distro the Windows reporter's address).
--portoninstall-serviceis saved for later runs. Discovery checks 8787–8790.--bind <address>pins the address;serveis an alias fordaemon. - Autostart:
install-servicekeeps the daemon running and restarts it if it exits. Linux gets a systemd user unit with lingering, so it starts at boot. Windows gets a hidden launcher in the per-user Run key that starts at sign-in, with no administrator rights needed. Inside WSL it registers a Windows sign-in launcher that starts the distro; WSL needs mirrored networking or Tailscale in the distro; each distro picks its own port.--printshows what would be installed;uninstall-serviceremoves it. - Check a machine:
burn-report doctor. - Limits (vendor quotas) need the tokscale CLI, which the reporter runs through Bun or Node if either is installed. Usage syncs without it.
| Tab | What you get |
|---|---|
| Overview | Spend or tokens, 24h to All. Drag the chart for any day. Agents, cache hits, limits. |
| Explore | Models, agents, workspaces, sessions. Ranked. |
| Systems | RAM, GPU, temps. 24h trends. Live fans and power mode on opted-in Linux machines. |
| Machines | Scan a QR code, tap the rest. Refresh or reconcile. |
| Settings | Theme, timezone, disconnect. |
- Pull down anywhere to refresh.
- Header pill shows progress: "Syncing 1/3".
- Numbers glide, and the total shows what the pull added.
- Incremental: only new rows travel.
- Warm: reporters scan in the background, so a pull does not wait for one.
- Resumable: big backfills continue while the app is open.
- Self-correcting: a daily reconcile catches late records and price fixes, downloading only from the first day that changed.
- Independent: a failed quota check never wipes good data.
- Health: sampled every 30s, 24h kept.
- Phone-only deletes: removing a machine clears the phone's copy, never the machine's.
- Machines: run the install line again. It replaces both binaries and restarts the reporter.
- Phone: install the new
burn.apkover the old one. History and machines stay. A copy you built yourself is signed with a different key, so Android asks you to uninstall it first. Uninstalling deletes the phone's copy of every machine and its history. Adding the machines back downloads what they still have, but history a machine no longer keeps (sessions an agent deleted or pruned) is gone for good, so uninstall only if you accept that. - From source: rebuild
burn-eventswith each pin bump; the reporter refuses one built for another tokscale pin. pushis gone: runinstall-serviceinstead. Services that still pass--no-liveor--live-urlexit with a hint, andinstall-servicedisables older hand-written Linux units that ran the reporter, keeping their files.- Kept: reporter identity, phone history, settings and machine URLs.
Deep dives: transfer behavior · why Tailscale-only
| Path | What |
|---|---|
apps/mobile |
Expo Android app |
apps/reporter |
burn-report CLI and machine backend (read-only, plus opt-in power mode) |
crates/burn-events |
Pinned Rust exporter of priced tokscale records (JSONL) |
packages/sync-api |
Transport, shared types, validation, mirror keys |
docs |
Decisions, transfer behavior, audits |
bun run typecheck
bun run test
bun run --cwd apps/mobile android # local Android build; needs JDK + Android SDK- Release:
scripts/release.sh 0.2.0bumps every version, runs the checks, tags and pushes. The tag builds the APK, both binaries for each platform and the installers into one GitHub release (how). - Rules: AGENTS.md
- Why Tailscale-only: ADR 0003
- Why the one write endpoint: ADR 0004
- Upstream
tokscale events --jsonl - Skia charts
- Accurate model-price reference transport



