AgentSpend
A macOS menu bar app that shows what your coding agents cost, split by agent, plus an estimate of the electricity behind it. Updated live as you work.
So "switch to a cheaper model" becomes a decision you can put a number on, instead of a hunch.
Download · Why I built it · The energy method
What it is
AgentSpend reads the logs your coding agents already write on your Mac and shows
what they cost, broken down by agent. Claude Code and Codex are read directly;
sixteen more agent CLIs are read through a bundled copy of
ccusage. It runs entirely on your Mac:
no API keys, no account, no telemetry. It picks up your usage automatically
and updates while you work.
Native Swift. No runtime, no framework, no Electron: a universal binary that
sits at 0% CPU when idle, alongside a copy of
ccusage for the agents it has no parser of
its own for. A tool that burns power to report power would be a bad joke.
Install
- Download
AgentSpend.dmg from the latest release.
- Open it, drag AgentSpend into Applications, and launch it.
That's it. The app is signed with an Apple Developer ID and notarized by Apple,
so it opens without any Gatekeeper warning.
It appears as a ⚡ menu bar item showing today's cost. There's no Dock icon
and no window. Click it for the panes below; the ⚡/$ button at the bottom of the
popover swaps the menu bar between cost and energy. Quit from there too.
What it shows
| Pane | |
|---|
| Home | Cost and energy over 1/14/30/90 days, today against your own daily average, then the one or two things worth acting on, limited to today so it speaks to the work still running |
| Sessions | Recent sessions: project, branch, model, cost, and context size, plus an hourly breakdown |
| Savings | The same analysis over your whole history: ranked, quantified changes to how you work, each with its evidence and its caveat |
| Method | Per-model baseline, every coefficient with its range and source, and the one slider that matters. Reached from the footer link |
How much to trust these numbers
The dollars are exact arithmetic on published rates. The energy is an
estimate, and it is shown for Anthropic models only.
If you are on a flat-rate plan (Claude Max, Pro, Team, a ChatGPT plan), the
dollar figure is what the same tokens would cost at API rates, not what you were
charged. The app reads your plan from Claude Code's own config and says so under
the headline, because the gap is large: a fortnight reading $2,272 costs about
$50 on Claude Max 5x. It is the right number for comparing one model against
another, and the wrong number to call a bill.
Energy for non-Anthropic models is withheld rather than guessed. Their tier
would have to be inferred from price, and price does not track compute across
vendors, so a figure would be directionally unreliable rather than merely
uncertain.
No AI company publishes how much power a token takes. So every energy figure
you'll see anywhere, including this one, is an estimate.
This estimate is built on the three best public measurements available (Google's
median Gemini prompt at 0.24 Wh, Microsoft's peer-reviewed 0.31 Wh, Epoch AI's
~0.3 Wh for GPT-4o), which land close enough to each other to be worth trusting.
Use the energy side to compare one model against another, not as a meter reading.
The Method pane shows every number the estimate uses, its range, and where it
came from, and lets you drag the single assumption that moves the total most. The
full reasoning is in the blog post.
Build from source
Requires a recent macOS and the Swift toolchain (no Xcode project needed).
cd mac
./build-app.sh # build build/AgentSpend.app (ad-hoc signed, runs locally)
open build/AgentSpend.app
Checks you can run:
./.build/release/AgentSpend --selftest # 158 assertions
./.build/release/AgentSpend --verify # aggregates, diffable against the Python oracle
./.build/release/AgentSpend --render <dir> # render each pane to PNG
./.build/release/AgentSpend --window # the panes in an ordinary window
tools/prototype.py is an independent Python implementation of the same parse
and model. It's the reference the Swift ingestor is checked against. On a frozen
snapshot the two agree exactly, to the token and the cent.
Repository layout
This is a monorepo:
mac/: the Swift menu bar app (see mac/README.md for
architecture, and mac/RELEASING.md for how releases are cut).src/, content/, public/: the Next.js landing site and blog.tools/: the Python reference implementation and verification scripts.
Privacy
Nothing about you leaves your machine. AgentSpend reads only your local
Claude Code logs and writes a small local database at
~/Library/Application Support/AgentSpend/. There is no telemetry, no
analytics, and no account.
It makes one network call. Once a day it asks the GitHub releases API for the
latest version number, and shows a link in the footer if yours is older. That
request sends no identifier and no usage data, so it can tell that somebody
checked but not who, or what they spent. It never downloads or installs
anything by itself. Switch it off with Check for new versions in the Method
pane and the app makes no network calls at all.
Not affiliated with any of these vendors
AgentSpend reads logs that coding agents write on your own machine, but it is an
independent project and is not affiliated with, endorsed by, or sponsored by
Anthropic, OpenAI, Google, GitHub, or any other vendor whose tools it reads.
Each agent's logo appears beside its share of your spend, to show whose usage a
figure represents. Those logos are the trademarks of their owners and are used
for identification only. See mac/NOTICE-logos.md.
License
MIT © 2026 Kapil Gowru
