Activity Ledger

A desktop app that reads your Claude Code transcripts on your own machine and shows where an agent harness spends its tokens: real work versus overhead, at API list prices. Inside the work, it shows the kind: coding, debugging, code review, research, and so on.

Activity Ledger is made by Pinkham Software. It is not made, endorsed or supported by Anthropic.

Everything stays on your machine. The app reads Claude Code's transcripts (~/.claude/projects, and those of any other Claude Code config folder, such as a second subscription's ~/.claude-work) and keeps its own cache in the app's data folder on this computer. It makes no network calls, except one optional call to a local model that you run yourself (see Privacy).

Figures are estimates at Anthropic's published API list prices. They are not an invoice, and they are not your plan's usage.

One row is not from the published list yet. claude-haiku-5-5 is priced by prompt size: $0.10 input and $0.50 output per million tokens for a prompt of up to 100,000 tokens (input plus cache writes plus cache reads), and $0.50 input and $2.50 output above that. These figures were supplied by the maintainer on 2026-10-15; check them against the provider's pricing page. No cache prices were given, so this model's cache reads and writes use the same multipliers as the other rows (0.1x input for a read, 1.25x or 2x for a write). That part is an assumption. The price editor shows only the lower rate; the rate above 100,000 tokens can be changed only by hand, with the app closed, in prices.json in the app's data folder (the file is written the first time you change a price or a pricing option).

Work vs overhead

• Overhead is what keeps the work going but is not the work: context compactions and, with a harness adapter on, its inbox handling, memory upkeep, standups and scheduled jobs.
• A segment a harness starts that goes on to do the work counts as that work, labelled by what it did.
• A harness is one adapter block in triggers.json; Munder Difflin is the first, off until you turn it on.

In this run the overhead share is small: most wake-ups went on to do real work, and standups were the trigger with the most upkeep.

What it shows

• Overview: spend, tokens and sessions over time.

• Cost and efficiency: cost per model, cache hit rate, and the price table the estimates use. You can edit the prices.

• Cost by activity: every session is cut into segments, one per prompt, and each segment gets a label. Cost, calls and tokens are totalled per label, per day, and split into work and overhead.

• Workflow, Projects, Sessions: tool use, cost per project, and a drill down into each session.

• Accounts: with more than one Claude subscription on the machine, every view can show one account or all of them (the picker in the header). With all of them, rows carry the account's colour and name, and the Overview splits the cost by account. See Accounts.

• Weekly report: one click saves a week (last week, or this week so far) as a Markdown file: totals against the week before, cost by day, model and project, and the five busiest sessions. It covers the account picked in the header (or all of them, and then names them). It is written only when you ask, holds no prompt text, session titles or paths, and can letter projects ("Project A") and number accounts ("Account 1") instead of naming them. The tray's "Save weekly report" opens the same save in the main window, for the account the ring follows.

• Export: a JSON summary with hashed ids (projects and accounts) and no prompt text, saved to a file you choose.

• Menu bar budget (optional, off by default): a ring in the menu bar for today's spend against one account's monthly budget, for anyone working to a company API budget. Each account has its own budget and thresholds. The budget is spread evenly over the calendar month, so a $300 budget in a 31 day month allows $9.68 a day; days start at local midnight. The ring is green below 80% of today's allowance, amber from 80% and red from 100% (both thresholds can be changed). A click opens today's spend, the month to date, the pace, the projected month end, what is left per day, and today's split between work and overhead, with a switcher at the top to follow another account. "All accounts" there shows their spend together, with no ring: there is no budget to measure it against. Turn the ring on, and pick the account it follows, under Cost by activity, Menu bar budget; set each account's budget under Cost by activity, Accounts. With it on, closing the main window keeps the app in the menu bar and out of the Dock.

  The figures are the day view's figures for the same calls, read every 60 seconds from what the transcripts gained since the last read. Both count this Mac's local days. Web searches are priced at $10 per 1,000 here, in the day view and in Cost by activity alike. Every figure is an estimate for this Mac's Claude Code only: not chat, not other machines, and not your invoice. If Claude Code's managed modelPricing setting is in /Library/Application Support/ClaudeCode/managed-settings.json, its rates are used and the panel says "at configured rates"; otherwise "at list price".

  Clicking the icon while the panel is open hides it on focus loss and the same click may open it again; close it by clicking elsewhere.

Accounts

Claude Code keeps one config folder per login: ~/.claude by default, or the folder CLAUDE_CONFIG_DIR names. A second subscription, for example a work one, is often run as CLAUDE_CONFIG_DIR=~/.claude-work claude. Activity Ledger reads every one of them:

• Found by itself: every folder in your home named .claude, .claude-* or .claude_* that holds a projects folder. On Windows the home folder of each WSL user is searched too.
• Grouped by login: each folder's account is read from the oauthAccount block of its .claude.json (for ~/.claude, the ~/.claude.json beside it): the account and organisation ids, the organisation's name and kind, and the email. Folders signed in to the same account and organisation are one account, so a Windows ~/.claude and a WSL ~/.claude signed in to the same plan count together. The same person in another organisation is another account. A folder with no login is an account of its own, named after the folder.
• Named for you: an organisation's name ("Acme"), or "Personal" for a Pro or Max plan, with the plan beside it. Each account gets its own colour.
• Set per account (Cost by activity, Accounts): its name, colour, monthly budget and thresholds, and whether it is hidden. A hidden account leaves the picker and every "All accounts" figure. Each config folder can be counted in another account, or split into one of its own. A folder outside your home can be added by hand. Refresh looks for new folders.
• Set for the whole app: everything else, for example the local model, the price table, the harness adapters, the report's hide setting, and whether the ring shows in the menu bar.

The transcripts do not say which login wrote them. A folder signed out and back in as someone else counts all its history in the login it has now; move it in the Accounts card if that is wrong.

How labels work

Rules first. Each segment is labelled by fixed rules over what it did, with no model call. The rules look at what started the segment (a prompt you typed, a compaction, a task notification) and at its tool mix: edits, notes, test runs and failing streaks, reads, web lookups, shell commands, and pull request or git range commands. The labels are:

Label	When
coding	it edited or wrote files (code, notes or docs)
debugging	it edited or wrote files, and 3 or more test runs failed in a row
code review	it read and ran tests, or ran pull request or git range commands, without editing source
research	at least 30% of its tools were web lookups
review / explore	at least 70% of its tools were reads
shell / ops	at least half of its tools were shell commands
chat / planning	it used no tools, or only planning tools such as a to-do list or launching a sub-agent
other work	none of the above
compaction	a context compaction
a harness trigger (inbox handling, memory upkeep, standup, scheduled job)	a harness adapter is on, its trigger started the segment, and the segment did only planning or calls on the harness's own files

Agent harnesses. triggers.json in the app's data folder lists adapters, one per harness. An adapter names the opening lines that start the harness's upkeep, the senders in its log that mark a standup or a scheduled job, and its bookkeeping files (inbox, memory, board). Calls on those files are bookkeeping, not work. Turn an adapter on or off on the Agent harness card; the app then reads the transcripts again. The By trigger table shows what started the work and how much of its cost was the work itself.

An optional local model second. If you run Ollama, the app can ask your model for a second label for each segment, from a fixed list. The model is given a one-line definition of each label and told to judge mostly by the tool counts. You start this yourself with Label now. It is off until you do.

The app asks the model for a JSON answer limited to the label list (Ollama's structured outputs) and turns thinking off, so a thinking model answers instead of reasoning. Use Ollama 0.9 or later: older versions cannot turn thinking off. Checked with llama2, qwen3:8b and gemma4:e2b. The first Label now loads the model, which can take minutes on a Mac without a GPU; the card says "Loading the model" meanwhile.

A toggle. On Cost by activity, switch between Tool labels and LLM labels. The totals are the same in both views. Segments the model has not labelled yet are shown as "not labelled by the model", so nothing is dropped or double counted.

Privacy

• The app reads Claude Code's transcripts in ~/.claude/projects (and each other config folder's projects) on this machine. It never changes them.
• To tell accounts apart it reads the oauthAccount block of each config folder's .claude.json: the account and organisation ids, the organisation's name and kind, and the email, shown only in the app. It never opens Claude Code's credentials file. The cache keeps the account id of each record; an export hashes it.
• Its cache (a SQLite file in the app's data folder) keeps counts, token totals, costs, timestamps, tool names, labels, model names, the Claude Code version, git branch names, message and tool call ids, the session titles Claude Code wrote, and the project folder and transcript file each record came from. With a harness adapter on, it also keeps each harness prompt's trigger label and the message ids the prompt named. It does not keep your prompts, Claude's replies, tool inputs or commands. Tool inputs are read in memory only to decide a tag such as "test run".
• The only network call is the optional labelling pass, to your Ollama. By default it must be on this machine (a loopback address). A remote host is used only if you tick "Allow a remote Ollama". For each segment it sends tool counts and a short excerpt of the opening prompt. Only the label that comes back is kept.
• The trigger file holds phrases and paths only. With an adapter on, the app reads that harness's log on this machine to pick a label (a standup, a scheduled job). It keeps the label and the ids of the messages a prompt named, nothing else from the log.
• The menu bar budget reads the same transcripts, and Claude Code's managed settings file for modelPricing only. It changes neither.
• No telemetry, no sign-in to Activity Ledger, no calls to Anthropic or anyone else.

Install

Download the signed app from the latest release: one zip that runs on Apple silicon and Intel Macs, signed with a Developer ID and notarized by Apple, so it opens like any other app. Unzip it and move Activity Ledger to Applications. To build it yourself instead, read on. On Windows, see Windows.

Requirements

• macOS or Windows 10/11 (other platforms are untested)
• Rust (stable)
• Node.js 22 or later
• On macOS: Xcode Command Line Tools: xcode-select --install
• On Windows: Microsoft C++ Build Tools ("Desktop development with C++") and WebView2, which Windows 10 and 11 already have

Build

npm ci
npm run tauri dev       # run in development
npm run tauri build     # build the app for this computer

For one app that runs on both Apple silicon and Intel Macs:

rustup target add aarch64-apple-darwin x86_64-apple-darwin
npm run build:universal

The built app is under src-tauri/target/.../bundle/macos/. A build you make yourself is not signed or notarized, so the first time you open it, right click it and choose Open. The download from the releases page needs no such step.

Windows

There is no signed Windows download yet; build it yourself.

Set up (once, in PowerShell; the Build Tools take a while and ask for admin):

winget install Microsoft.VisualStudio.2022.BuildTools --override "--quiet --wait --norestart --add Microsoft.VisualStudio.Workload.VCTools --includeRecommended"
winget install Rustlang.Rustup
winget install OpenJS.NodeJS.LTS
winget install Git.Git

WebView2 is already part of Windows 10 and 11. Keep the repository on a Windows drive (for example C:\projects\activity-ledger), not inside WSL: npm and cargo are slow and unreliable over \\wsl.localhost.

Build and run with the same commands as above. npm run tauri dev runs the current code without installing anything. npm run tauri build makes an MSI and a setup .exe under src-tauri\target\release\bundle\msi\ and bundle\nsis\.

• Smart App Control must be off to build: every Rust build runs small programs it compiles itself, and Smart App Control blocks them. It has no per-app exceptions. It also blocks running an unsigned build. Turn it off under Windows Security, App & browser control, Smart App Control.
• SmartScreen warns the first time you run the unsigned installer; choose More info, then Run anyway.
• Reinstalling the same version does nothing: the MSI sees the app as already installed. Uninstall it first (Settings, Apps; your data in the app's data folder is kept), or raise the version.

What is different on Windows:

• Transcripts are read from %USERPROFILE%\.claude\projects and, for Claude Code run inside WSL, from each WSL user's config folders in each distro (~/.claude, ~/.claude-work and the like; distros found with wsl.exe -l -q, Docker Desktop's skipped). Reading a distro that is not running starts it. Press Refresh to read what is new. The accounts card says where each folder was found ("Windows", "WSL · Ubuntu").
• The budget ring sits in the system tray, and its panel opens above the taskbar. Windows 11 puts a new tray icon in the overflow (the ^ next to the clock); drag it out to keep it in view.
• Claude Code's managed modelPricing is read from C:\Program Files\ClaudeCode\managed-settings.json.
• The main window opens maximized (src-tauri/tauri.windows.conf.json).

CI builds and tests macOS and Windows on every pull request and push to main (.github/workflows/build.yml).

Signed build (maintainers)

scripts/sign-notarize.sh builds the universal app, signs it with a Developer ID and the hardened runtime, has Apple notarize it, staples the ticket and zips it. It runs on the Mac that holds the certificate, uses the identity and the keychain profile by name only, and reads no secret from a file or a variable. The comments at its top list the one-time setup.

Test

(cd src-tauri && cargo test)
npm run test:ui
npx tsc --noEmit

199 Rust tests and 42 front-end tests; on Windows the 4 tests of the macOS signing script are skipped. The weekly report's expected output is checked in at src-tauri/tests/fixtures/weekly-report.golden.md; after a deliberate change, rewrite it with UPDATE_GOLDEN=1 cargo test report:: and read the diff.

License

MIT. See LICENSE. Third-party licenses are in THIRD_PARTY_NOTICES.md; regenerate that file with python3 scripts/third_party_notices.py > THIRD_PARTY_NOTICES.md.

Claude and Claude Code are trademarks of Anthropic. The MIT license covers this code, not those names.

Questions, bugs or ideas: open an issue on this repository.