Pi and OMP session files are the only usage ledger there is
Local-first tools put the record on your disk. Three ordinary events erase it.
·2 min read
Pi and Oh My Pi are local-first coding agents, which means the usage record is a folder on your disk rather than a page on a vendor site. That has a measurable consequence for anyone comparing tools: the number you need exists, but only if something reads it before the folder disappears. The UsageAtlas guide covers both formats.
The two session formats
| Tool | Format | What a reader gets |
|---|---|---|
| Pi | Local session files | Model, tokens, timestamps |
| Oh My Pi | Local session files in OMP format | The same fields |
| Both | Per-project folders | Working directory when recorded |
UsageAtlas reads both, which allows a comparison that no single vendor can offer: the same week of work, two tools, one view.
Why this matters more than a dashboard
A hosted tool can tell you what you spent last month because it kept the record. A local-first tool hands that responsibility to you.
Three failures follow if nobody takes it.
The history is lost on a rebuild. A container or a wiped home directory takes the sessions with it.
Two machines split the record. Half the sessions on a laptop and half on a desktop produce two partial truths.
Sharing transcripts instead of aggregates. The first is a privacy problem, the second is a number.
The fixes are small: back the folder up with dotfiles, nominate one machine per project, and export counts rather than conversations.
Setting up the read
Run Pi or OMP so sessions exist, install the tracker, and confirm it detected the session directory. Today gives the live session and model; History gives 7, 30 or 90 days.
If detection fails, point the app at the directory. Non-default locations are common with tools that let you move state.
Reading tokens per session rather than per week
This is the comparison trap. A tool used twice a week will look cheap in a weekly total even if each session is expensive.
Divide by sessions instead. That number answers "which tool costs more per unit of work", which is the question a routing decision actually needs.
The context effect
Input tokens scale with thread length, so a single long session can dominate a week without any unusual work. When the history shows that pattern, the fix is a new thread rather than a new subscription.
Output tokens tell a different story: they track the size of generated changes, so a refactor will always look heavier than a question.
What the record cannot carry
Allowances. Session files hold usage, and entitlement lives with the provider. Any local tracker that implies otherwise is guessing.
FAQ
Does Pi write usage data automatically?
Yes. Sessions are recorded as local files with model and token information per turn.
Can one view compare Pi and OMP?
Yes. Both are supported providers, and History shows them side by side.
Is cloud sync required?
No. Local tracking works without an account, and sync is optional.
The takeaway
Local-first tools trade a vendor dashboard for your own responsibility. Read the files, back them up, and compare per session rather than per week.