Kamol Nazarov The back office lab Brooklyn, NY

N°06 · Open-source tool · 2026 · Public beta

BurnRate

See what your AI coding would cost at API prices, and which subscription runs dry first, before it does.

Client
Open source, Apache-2.0
Role
Author
Built with
Python 3.12, FastAPI, SQLite (WAL), vanilla JS, Tauri v2
Status
Public beta · v0.3.0-beta.15
Live at
github.com/kamol-nazarov/myBurnRateOfficial
A monitor at night showing the BurnRate dashboard with the Sidecar HUD pinned on top.
Demo data

What it does

  1. 01

    Reads, never intercepts

    Eight AI coding tools are read from the logs they already keep, opened read-only. BurnRate never sits on the inference path.

  2. 02

    Run-dry forecasts

    Every plan window gets a meter, an even-pace tick and a reset countdown, so “tight” shows up while there’s still time to change course. A second Claude Code account gets its own lane (experimental).

  3. 03

    Refuses to invent numbers

    Unpriced usage shows “—”, never $0. The cache-hit rate is withheld when telemetry is incomplete.

The thing to see

The Sidecar HUD as two polls land: spend, tokens and plan meters update. Synthetic demo data from the repo’s fixture.

Proof

  • 8AI coding tools read from local filesREADME + adapters
  • 948test functions at v0.3.0-beta.15tests_spend/
  • 17versions in 46 days, Sep 4 to Oct 20, 2026CHANGELOG

No savings figure is claimed. Most dollar values are published-rate estimates, and the UI marks them “≈”.

On screen

The BurnRate dashboard: “What runs out first”, with Claude Code at 40%, a second account, Claude Code · Work, at 27%, and Codex at 25%.
The Sidecar HUD in its graphite and paper themes, with a lane for each plan, the Work account included.
The dashboard in the paper theme.
Detail of the plan lanes: each account’s 5-hour and weekly meters, an even-pace tick and the time to reset.

Rendered from v0.3.0-beta.15 fed the repo’s synthetic fixture plus an invented second account. No real usage was read.