See what your AI coding costs, by project and by commit.
TokenSpend AI reads the usage logs your coding assistants already write and turns them into spend per project, task and commit, with budgets and forecasts. It runs on your machine. There is no account to create and nothing to send anywhere.
Meet the Creator: Jonathan Davidar ↗ (opens davidar.org in a new tab)
I solve problems with creative vision and AI solutions architecture.I turn open-ended needs into designs people can use and systems teams can run.
- GitHub Copilot$16941%
- Claude Code$11127%
- Codex$5814%
- Qwen Code$4511%
- Continue$297%
69% of the $600 budget · 55% of the month gone
376K of 480K input tokens from cache · Healthy
$412 over 145 commits
- Plan the task
Set an hour estimate, a labor rate and whether the work is billable.
- Start its timer
Work as usual. Requests in that project during the interval join the task.
- Stop, review, export
Reviewed billable hours become fees. Export the timesheet as CSV or put it in the management report.
| Task | Estimate | Tracked | Time left | AI spend |
|---|---|---|---|---|
| Implement sign-in validationcheckout-service · Billable | 6.00h | 5.25h | 0.75h left | $18.40 |
| Refactor billing exportcheckout-service · Billable | 4.00h | 4.75h | Over 0.75h | $31.10 |
| Write migration testscheckout-service · Nonbillable | 3.00h | 1.50h | 1.50h left | $6.85 |
Attribution is by time. If two work intervals overlap, those requests stay unassigned instead of being guessed. Entries start as drafts, and only reviewed hours count toward fees. AI spend per task uses the best figure available for each request: GitHub’s bill where it is matched, then GitHub’s per-call meter, then the estimate.
For developers
Nothing to configure. It finds the logs that are already on your disk.
- Works with the tools you useGitHub Copilot (VS Code and the Copilot CLI), Claude Code, Qwen Code, Codex and Continue. Each is labelled, and none is mistaken for another.
- Cost where decisions get madeSpend by project, repository, task and commit, a month-end forecast, and dials for budget burn, cache health and context pressure.
- Advice you can checkSpots light requests sent to expensive models and shows the monthly saving. It cannot judge answer quality, so you test before you switch.
For IT and finance teams
Built so you can approve it without a long review.
- Local by designRuns on the developer’s machine. No TokenSpend AI server, no account, no telemetry. The dashboard listens on 127.0.0.1 only, and every request for your data needs a private key.
- Metadata onlyStores the model, time, token counts, project and cost. Never prompts, replies, code or tool output. The test suite checks this.
- Small network footprintPublic price lists are fetched with plain requests that send no data. If a developer chooses to connect GitHub billing, it uses a read-only token and talks only to api.github.com.
- Reports you can hand overBudgets per project, executive and management reports that print to PDF, and CSV and JSON exports.
What it does not do yet
So nobody finds out later.
- Each install reports for its own machine. There is no team-wide rollup yet.
- Copilot usage paid by an organisation cannot be read with a personal token, because GitHub limits organisation billing to its administrators. The local meter still covers Copilot CLI use.
- Local logs cannot tell GitHub accounts apart, so billing connects one account at a time.
- Verified on Windows. macOS and Linux have not been checked yet, and the security hardening has not had an independent review.