How much of what they wrote is still in your code a week later, line by line — next to what the session cost.
pipx install aiwatcher-local · reads Claude Code, Cursor, Codex, git · no account
Running AI inside your product? See AIWatcher Enterprise →
Twenty-seven commits from this repo. Switch the measure and watch every one of them agree with every other.
Pick a bar to see which commit it is, what it cost, and how much of it is still in the repo.
A revert adds a new commit. Deleting every file it created changes nothing about the old one. Anything merged answers yes forever.
Both directions, because a commit whose whole job was deleting dead code would otherwise score zero — exactly backwards.
A measure that answers yes to everything can’t rank anything. Resolution is the product, not a lower number.
Every commit gets a cost, and a count of what’s still there. Put them together and the spread is enormous — and invisible to anything that stops at the invoice.
27 measured commits from this repo’s own history. Cost is API-equivalent, not a subscription bill. These counts are a floor.
No account, no key, no config. AIWatcher reads the agent history and git log already sitting on your machine, so there is nothing to start collecting.
Measuring what is still there needs a commit to be a week old, so today’s work reads as not yet measurable — never as zero.
All of it read from the agent history and git log already on your machine. No wrappers, no SDK, no code changes, nothing sent anywhere by default.
You already have a tool that knows exactly what you spent, and a tool that knows exactly what is in your repo. They have never been introduced.
Tokens, dollars, model, date. To the cent.
Whether a single token of it ever became code that stayed. It has never seen your repo.
Every line, who committed it, and whether it is still there.
Which agent session wrote it, how many turns it took, or what it cost. Your name is on all of it.
One side knows the spend, the other knows the code. Joined, they answer the question neither can: what did the money actually produce?
AIWatcher replaces neither of them. It reads both, on your machine, and closes the gap between them.
Apache-2.0, source on GitHub. Every number on this page came out of a local file that you have too.
Files these tools already wrote to your disk. Nothing new is instrumented.
One file you can open, grep or delete. Raw prompt text from Prompt Gate is never written down.
The network stays off unless you run an update command, switch on update checks, or configure a connected workflow.
Optional AI Assist is the one exception: it sends bounded context, only after you turn it on and pick a workflow that uses it.
Once agents write a real share of the work, the bottleneck moves to review. AIWatcher Enterprise answers one question on the pull request: does this change carry enough trustworthy evidence to review efficiently?
See AIWatcher Enterprise →Well-evidenced work stays quiet. Exceptions carry one reason and one action — and never a green check the evidence does not support.
Local is where that evidence starts. It leaves as a minimized export you preview and control, never as your prompts or your source.
no account · no API key · works offline
Whatever we build for organizations is built on top of this, never by taking it away. Raw prompts and source stay on your machine by default. Anything an organization ever sees is an explicit, minimized export you can preview, disable and revoke.
Free and open source, forever · no green check we can’t back
Free, open source, no account, nothing leaves your machine by default.
Install AIWatcher LocalWe’re taking on a small number of design partners for the organization-wide loop.
Tell us about your setupWe reply within 48 hours. No newsletter, no marketing automation. Just a real reply from one of us.