Your agents spent $4,054. On what?

A cost ledger for teams that use AI coding agents. A small collector on each developer's machine reads the session logs Claude Code and Codex CLI already write, uploads per-session metadata (tokens by model, repo, branch, developer, timing), and Tallyhook prices each session at the vendor's public list rate and attributes it to a repo by its git remote. You then group repos the way your business works: by team or service to see where the budget goes, or by client to bill it back with your markup.

Engineering teams and agencies of roughly 2 to 50 developers using Claude Code or Codex. In-house teams use it to answer which repo, squad or service the spend belongs to, whether it is accelerating, and whether any single session ran away. Agencies and consultancies use the same ledger to bill the cost back to the client it belongs to, with a markup. One product, and the grouping is the only difference.

No, and it is not built to. The default views are per repo and per team, because that is where a spending decision actually gets made. Per-developer totals exist, since you need them to size seats and spot a runaway loop, but they are never the headline, nothing in the product scores or ranks people, and a read-only report shared outside the workspace leaves the per-developer table out entirely. If you are looking for a tool to performance-manage engineers by AI usage, this is the wrong one.

It is a list-price equivalent: what that usage would have cost on the API, priced from the vendors' public rates. On a subscription that is the value of what your team used, not an invoice you received, and the app says so wherever a number appears. It is the only way to get a dollar figure at all for subscription usage, which vendor dashboards leave blank.

Vendor dashboards report usage per user and per model. None of them know which repo a session belonged to -- let alone which team or client that repo belongs to -- and usage inside a subscription seat has no dollar figure at all. Tallyhook attributes every session to a repo by its git remote, rolls repos up into whatever groups you use, prices it at list rates, and exports a cost report or invoice lines.

ccusage is an excellent open-source CLI for one machine and was the reference for token counting. Tallyhook is the team layer on top of the same local logs: every developer's sessions in one place, grouped by team or client, with cost alerts, per-group budgets, a month-end forecast, shareable read-only reports and invoice output. If you only need your own machine's numbers, ccusage is enough -- and so is our own free npx tallyhook. The paid product is for when the question is about a team rather than a laptop.

Tokens come from the vendor's own session logs, deduplicated per message (including messages that forked subagents copy from the parent session). Prices are the public API rates per model, including cache reads, 5-minute cache writes at 1.25x input and 1-hour cache writes at 2x input. Claude fast mode is priced at its premium rate. You can override any model's price for negotiated rates. It is a list-price equivalent, not a figure reconciled against your vendor invoice (on subscription plans there is no per-session invoice to reconcile against), and it leaves out server-tool fees such as Claude web search, the 1.1x rate for US-only inference and OpenAI long-context surcharges.

Yes. The collector is MIT-licensed, one dependency-free Node.js file, with its source on GitHub (github.com/tallyhook/tallyhook-collector) and published on npm as tallyhook, so npx tallyhook and npx tallyhook install <token> both work. You can read exactly what leaves your machine before you run it. The hosted web app and API are not open source.

Only metadata: token counts by model, start and end time, tool and version, git remote and branch, which repositories the edited files belong to, developer identity from git config (or the OS username if no git email is set), the machine hostname, turn and tool-call counts, paths of edited files, and optionally the first 160 characters of the first prompt. Transcripts, code, diffs, tool output and environment variables never leave the machine.

Only developers who install the collector and upload sessions. Anyone else you invite, such as a project manager or account owner, signs in free as a workspace member (up to 50 people per workspace). The free trial covers whichever plan you are trying — 5 developers on Starter, 20 on Standard — and you can switch between them free in Settings while the trial runs, so you can evaluate with your real team before paying.

Add extra seats to the Standard plan any time from Settings: $7 per seat per month, up to 500 developers in total. That is below Standard's own blended rate ($149 / 20 = about $7.45), so growing past 20 never costs more per head.

Rarely, and briefly. The Stop hook syncs at most once every ten minutes and usually finishes in well under a second. Each upload times out after eight seconds, Claude Code caps the hook at 20 seconds, and the hook always exits cleanly, so it never fails a turn.

No. The Starter plan at $49 a month works for a single developer, and a solo consultant is the clearest case for it: you are the one eating the agent cost and the one sending the invoice, so there is nobody to roll anything out to. You install the collector on your own machine and get the invoice document with your markup, a read-only link to send each client, and history that outlives the session logs Claude Code eventually deletes. The free npx tallyhook command already groups repos into clients locally; the paid plan is what makes that into something a client can open.

Yes, two ways. The live demo at tallyhook.dev/demo opens the real app on sample data for a made-up five-developer agency: overview, clients, budgets, sessions and a finished invoice. Or run the collector itself (npx tallyhook, an open-source npm package) to print what each repo on your own machine cost over the last 30 days at list price, and npx tallyhook --by team to roll those repos up into your squads and services — or --by client, the same grouping with your markup on top, if you are billing it. It needs no account or token and uploads nothing; it only downloads the public price list.

Yes. Add a webhook URL in Settings and both alerts are POSTed there as well as emailed. Discord is verified: a real cost-spike and a real budget alert were both watched arriving in a channel, correctly formatted. The same payload also carries the field Slack documents for incoming webhooks, so it should work there too, but nobody has watched one arrive yet and we would rather say that than imply it is tested. The message is deliberately plain text, because Discord and Slack use different bold syntax and either one would render as stray asterisks in the other. Saving a URL sends a test immediately.

Yes. The workspace has a default markup and any client can override it on its own page; an empty field means it inherits. If one invoice covers several clients at different rates, each client's line is priced at its own rate and the per-developer and per-model billable columns come off, because no single markup describes them honestly.

Yours, once you set it. Workspace owners pick a billing timezone in Settings and every calendar-month range is cut on that clock. It matters more than it sounds: a session run at 8pm on the last day of the month in US Pacific is already the next day in UTC, so without a timezone it would be billed in the following month. Sessions are always stored in UTC and nothing is repriced; only the period boundaries move. The default is UTC.

Yes. Put a monthly budget on any client and workspace owners get one email when that client's month passes 80% of it and another at 100%. Separately, a per-session threshold emails you the moment a single session gets expensive, which is how you catch an agent loop the same day.

Cursor does not write per-request token usage to a local log the collector can read, so it needs a proxy, which is a bigger trust ask than reading a log file. It is not built. There is a one-line way to register interest at the bottom of the features page, and what gets built next comes from those asks rather than a guess.

Yes. Run npx tallyhook mcp and you have a Model Context Protocol server over stdio; in Claude Code, claude mcp add tallyhook -- npx -y tallyhook mcp. It exposes four tools: spend grouped by repository, the same grouped by model, the individual sessions that cost the most, and a summary of this period against the equally long one before it. It needs no account and no token and uploads nothing, because it reads the same local logs the free report does. If you have set "privacy": true, it leaves the prompt snippet out too.

Yes. Set "privacy": true in ~/.tallyhook/config.json and only counts and identifiers are sent.

Nothing is deleted and the collector keeps recording, so your history has no gap. The ledger itself (overview, clients, sessions, invoices, exports and client report links) is paused until you pick a plan, and opens again the moment you do.

Yes, any time, from Settings. Plan switches are prorated with no double charge; cancelling stops billing but never deletes your data. Workspace owners manage billing.