Understand your AgentSpend
A Mac menu bar app that shows you what your agent usage costs in dollars ($) and energy (kWh). It updates while you work and proactively teaches you habits to reduce your spend.
Free and MIT licensed.
Daily spend
82% of your spend ($445) is on top-tier models. Sonnet 5 is roughly half the energy and a third of the price per token.
opus-4-8 cost you $445 across 1,046 requests. The same tokens on sonnet-5 would be $91.
Averaged 118k of prompt per request across 214 requests, against a 34.2k median across your sessions.
$174 of $542 (32%) and 8.4 kWh.
A ceiling, not a plan. Nobody should run everything on one model, and a weaker model that needs more turns gives the saving back.
Suggested savings
Session spend
Why you can trust it
No network calls
It never phones home. It reads your local logs and nothing leaves your machine.
No API key, no account
Nothing to configure. It finds the logs your agents already write and picks them up automatically.
Stays on your Mac
The only thing it writes is a small database in your Application Support folder.
0% CPU when idle
Native Swift with no external dependencies. A tool that burned power to report power would defeat itself.
Open source, MIT
Every source, coefficient and caveat ships as one JSON file inside the app, meant to be read and argued with.
Checked against a second implementation
A separate Python version of the same parser and model is the reference the Swift one is tested against. On a frozen snapshot they agree to the token and the cent.
Questions
Do I need an API key or an account?
No. It reads the logs your agents already write on your Mac. No keys, no account, no network calls.
I'm on a Pro or Max subscription. Will the dollars match my bill?
No, and the app says so. On a flat rate plan it reads your tier from Claude Code's own config and labels the figure list price, not your bill. It is the right number for comparing one model against another, and the wrong one to call an invoice.
Does it work on Intel Macs?
Yes. The release is a universal binary covering Apple Silicon and Intel, about 7 MB. It needs macOS 14 or later.
Where does my data go?
Nowhere. It makes no network calls beyond a daily version check you can switch off. It reads your local agent logs and writes a small database in your Application Support folder.
Why is the energy a range instead of a number?
Because a truthful energy figure is a range. Nobody publishes per-token energy for these models, so every figure out there is a model rather than a measurement. AgentSpend shows you every coefficient, its range and its source, and lets you drag the assumption that moves the total most.
Which agents does it track?
Claude Code and Codex are read directly. Sixteen more agent CLIs are read through a bundled copy of ccusage, including Gemini, GitHub Copilot, OpenCode, Amp, Droid, Goose and Qwen. Each appears as soon as you use it. Cursor is not supported: it bills per request on the server and its local token data is too incomplete to trust.
How do I uninstall it?
Drag the app to the Trash. To remove it completely, delete its folder in Application Support too.
Is this an Anthropic product?
No. It reads logs these tools write locally, but it is an independent project with no affiliation to Anthropic, OpenAI, Google, GitHub or any other vendor it reads.