Tokenflow - See where your AI spend actually goes

Hey PH, Vimox here 👋 My own landing page once said a 102-line bug fix cost $2,499 in AI tokens. It cost $16. The other $2,483 came from the same branch over the next six days — a stale checkout that kept the branch name after the PR had already merged. No console showed me the difference. Mine didn't either. So I published the wrong number. That mistake became rule #2 of TokenFlow: A pull request owns the AI turns up to its merge — and nothing after it. It also convinced me that I didn't need another AI cost dashboard. I needed cost controls closer to where the code is actually being written. ### What TokenFlow does Every push attaches a cost receipt to the commit as a Git note. A GitHub Action turns that into a PR comment + status check, so a change that exceeds its declared budget can fail before merge: ```yaml - uses: vimoxshah/tokenflow@v1.3.5 with: max-usd: '50' ``` There’s also: → A guard hook that prices the live agent session and warns or blocks when it hits your cap. → MCP support so the agent itself can see what it has spent while working. → Ticket-level costs that connect branches referencing the same Jira, Linear, or GitHub issue. → FOCUS-shaped exports that let you put estimated AI spend next to your actual cloud bill. ### What 9 weeks of my own agent logs taught me I analyzed $26.7K of estimated agent spend: * 86% went to re-sending context rather than fresh work. * 93% of spend came from the top 1% of sessions. * The median session cost just $0.02. * After roughly turn 50, a turn became ~15× more expensive. That last number matters. That's where the brake belongs. One developer. Nine weeks. Estimated using published list prices. You can reproduce the analysis on your own logs: ```bash tokenflow receipt --sessions ``` ### A few things I'm deliberately being honest about TokenFlow is about honest numbers, so: * Costs are estimates at published list prices, not invoices. Unpriced models show as `unknown`, never `$0`. * Missing receipts never block your workflow. If no hook is installed, the Action logs one line and exits 0. * The self-hosted GitHub App ships and is documented, but I haven't personally exercised it end-to-end against a live App yet. * The Codex guard can warn but cannot block because Codex doesn't expose a blocking hook. * The macOS app is unsigned because I don't have an Apple Developer account. First launch requires Open Anyway. Free. MIT licensed. Zero runtime dependencies. `npm install @vimoxshah/tokenflow` adds exactly one package: itself. 841 tests. Green on macOS, Linux, and Windows across Node 22 and 24. ```bash brew tap vimoxshah/tap https://github.com/vimoxshah/tok... brew install --cask tokenflow npx @vimoxshah/tokenflow@latest setup ``` Want to try it without installing anything? Run TokenFlow in the browser with synthetic data: https://vimoxshah.github.io/toke... I'll be in the comments all day. What would make AI cost control genuinely useful in your workflow?