AI Tools
Muse Code tokens: read input, output and session totals separately
A weekly total blends a long thread with a quiet day and describes neither.
Muse Code tokens are recorded per turn in local session files, which makes them readable without an account, an API key or a network connection. The useful question is not whether the data exists but which number answers the decision in front of you. The UsageAtlas guide covers the file locations and setup.
Three numbers, three uses
| Number | What it tells you | When it helps |
|---|---|---|
| Input tokens | How much context each turn carried | Deciding to start a new thread |
| Output tokens | How much the model generated | Comparing a refactor with a question |
| Session total | What one piece of work cost | Spotting the expensive exception |
| Model split | Where volume went | Routing routine work |
A single weekly total answers none of those well. It blends a day of small edits with one long agent run and reports an average that describes neither.
Why input grows and output does not
Context is re-sent on every turn. Add a file, and every later request in that thread carries it. That is why the tenth edit in a long conversation costs more than the first even when the edit itself is identical.
Output tracks the size of the change. A rename produces little; a generated module produces a lot.
Taken together, the two explain the most common surprise in a tracker: a week with few sessions and a high total, which is almost always one thread that grew.
Setting up the read
Run Muse Code so sessions exist, install the tracker, and let it detect the data directory. Today shows the current session; History covers 7, 30 and 90 days.
If the app finds nothing, check the folder. Muse may be writing to a non-default location, and pointing the app there is faster than reinstalling anything.
The limit question
Local files do not contain plan allowances. A tracker can report what you spent and cannot report what remains, and the honest ones say so rather than drawing a meter from nothing.
That leaves the provider page as the place to check before a deadline, and the local history as the place to change behaviour before you get there.
Per project, when the record allows
Working directories in the session files let the app group by repository. That is the number a lead actually wants: which project consumes the budget.
When the grouping is absent, it means the field was not recorded. It is not an error to fix.
A two-minute weekly review
Open History, note the model split, find the largest session, and write one line. If the heaviest model is doing routine work, move it. If one session dominates, split the thread next time.
That is the whole practice, and it costs less than reading a dashboard every morning.
FAQ
Do I need an account to read Muse Code tokens?
No. Local tracking needs no account, and prompts are not uploaded.
Can the tracker show a plan limit?
No. Session files do not carry allowance information.
Is the editor status bar enough?
It shows current context use, not history, so it cannot answer what last week cost.
The takeaway
Read the tokens that match the decision. Input for thread hygiene, model split for routing, session total for exceptions.