pandev

What your AI coding agents actually cost — broken down by task, branch, model and file.

Runs on your machine. No account, no daemon, no network.

npx pandev@beta

pandev-metrics.com/cli · по-русски · How to verify the privacy claim · Changelog

Beta is live

npx pandev@beta runs the current beta on macOS (Apple Silicon and Intel) and Linux x64 — no install step, nothing else to set up. The Windows npm package is not published yet; on Windows use the beta installer from PowerShell: iwr https://raw.githubusercontent.com/pandev-metriks/homebrew-pandev-cli-beta/main/install-experimental.ps1 -UseBasicParsing | iex (the command there is pandev cost).

Beta means the numbers are already trustworthy — the parsers are verified against live logs — but command names and output may still change between releases.

---

---

The problem

Your agent bills you per token. Your work is organised in tasks. Nothing connects the two.

Every tool in this space reports sessions: "yesterday you spent $61." Useful once. It never answers the question you actually have — which piece of work was expensive, and why.

Which task cost what

That is the table above, and it is the whole point of this tool.

Note its last line. Work that happened outside a repository, or on a branch with no ticket key in its name, cannot be attributed — so it is reported as its own number instead of being spread across the tasks to make the table look complete. pandev task breaks that remainder down by project, so you can see what it consists of rather than guessing.

Getting this right is harder than it looks. Claude Code writes HEAD into its logs instead of a branch name — every single event, no exceptions. Per-task numbers built naively from those logs are wrong, and wrong by a lot: on our own data a single task came out overstated 70×. pandev reconstructs the real branch at each timestamp from git reflog. That reconstruction is what makes the number above worth reading.

Then: which prompt inside that task

Pick the expensive one and look inside it.

One prompt took 38% of the whole task. That is the actionable unit — not the day, not the session, but the specific thing you asked for and what it cost to answer.

You also get what the totals hide: active time against wall-clock, which model actually did the work, which editor it ran in, and how many rounds each file went through.

Prompt text is read straight from your local logs to print this. It is never stored and never sent — see Privacy.

And underneath: where the volume goes

Cache reads, cache writes, fresh input and output are four different prices. A bill that looks alarming is usually cache reads — and the fix for that is a prompt change, not less work. You cannot see any of that from a single total.

If you are on a subscription rather than API billing, the same view tells you what your usage would have cost at API rates, which is the only honest way to know whether the plan pays for itself.

Commands

npx pandev@beta	summary for the last 14 days
npx pandev@beta today	today only
npx pandev@beta task	cost per task
npx pandev@beta task WEB-812	one task: every prompt, time spent, files touched, tools used
npx pandev@beta files	cost by file, with edit rounds
npx pandev@beta models	cost and cache behaviour per model
npx pandev@beta why cache	how prompt caching actually played out
npx pandev@beta why ratio	context read per unit of output
npx pandev@beta web	build and open the dashboard
npx pandev@beta privacy	what is read, and what leaves this machine

Add --json to any command for machine-readable output, --days N to change the window.

Privacy

This tool reads your agent logs. Those logs contain your prompts, fragments of your code and sometimes your secrets. You should not take anyone's word about what happens to them — including ours.

So, plainly:

• No network code. The binary opens no sockets, contacts no host, and carries no telemetry, no version check and no analytics of any kind.
• No account. There is nothing to sign into.
• Nothing is written except the dashboard file you explicitly ask for with pandev web — created 0600, owner-only, in your own home directory. It contains your prompt text, so treat it like the logs it came from.

npx pandev@beta privacy prints the full list of what is read and what is taken from it.

And here is how to check that for yourself, without trusting us — network activity is observable from outside the process, so the claim is verifiable even though this is not open source: docs/VERIFY.md.

Requirements

Node 20 or newer for npx. Optional but recommended: git on your PATH — without it, per-task attribution is switched off and you get totals only.

Reads logs from:

• Claude Code — ~/.claude/projects
• Codex CLI — ~/.codex/sessions
• opencode — ~/.local/share/opencode/opencode.db, read-only via your local sqlite3
• ZCode — ~/.zcode/cli/db/db.sqlite, read-only via your local sqlite3

Nothing else is scanned.

Working in a team?

This tool is deliberately single-player. It reads what is on your machine and answers questions about your work.

If the question turned into "how does this compare across the team", "which projects burn the most", or "what does AI actually cost us per delivered ticket" — that is PanDev Metrics, and it is a different product.

License

Proprietary — see LICENSE. Free to install and run on machines you own or control, including at work. Redistribution, modification and derivative works are not permitted.

Cost figures are estimates computed from published provider rate tables. They are not an invoice. Reconcile against your provider's own billing.

---

_{Built by PanDev · pandev-metrics.com/cli · Found a bug? Open an issue}