tokenflex/ing
Manifesto

We measure what you burn, so you can flex it.

TokenFlexing is the public leaderboard for developer AI usage. This page is the why behind that pitch — what we promise, what we refuse to do, and how we plan to stay honest about both.

Tenets · in order of importance

  1. 01

    Developers should be able to brag in public.

    Token burn is the new GitHub commit graph — proof that you ship with AI. Right now it's locked inside vendor dashboards no one else can see. TokenFlexing turns it into a public, comparable, shareable number.

  2. 02

    Local files first. Cloud sync second.

    The CLI reads session JSONLs on your machine — Claude Code, Codex CLI, OpenCode — and never sends prompts or completions anywhere. Only the usage blocks. You can run `tokenflexing stats` offline forever and never pair with the cloud.

  3. 03

    Admin keys are a fallback, not the path.

    Most provider Admin APIs (Anthropic, OpenAI) require org-owner credentials that the average developer doesn't have. The CLI ships first because anyone with a Pro / Max / Team subscription can use it. Admin keys exist for the people who actually have them.

  4. 04

    Subscription tools deserve real numbers.

    Cursor and Claude Desktop don't write per-message tokens to disk — vendors hide subscription usage behind dashboards. We work around that with three sync paths for Cursor (cookie API, Downloads watcher, one-time CSV) and a Claude Code bridge for Claude Desktop. No estimates we don't trust.

  5. 05

    Privacy is non-negotiable.

    Every cloud-synced row is opt-in. Public, aggregate-only, or fully private — your call. Off-leaderboard is a single toggle. We never sell or share data with third parties. Your tokens, your call.

  6. 06

    The CLI is the source of truth.

    Anything the web app shows comes from `tokenflexing sync` or a paste-able API key. No proprietary client in the middle. If we ever shut down the web app, your CLI + local files keep working. Open-source CLI is on the roadmap once the schema is stable.

  7. 07

    Agents should run setup, not type it.

    TokenFlexing ships an MCP server with read tools (your stats inline in Claude / Cursor / Codex) AND action tools (the agent can pair your device, install the daemon, import a CSV without you opening a terminal). The future is your assistant doing the work, not handing you commands to copy.

Hard nos

What we won't do.

  • Sell, share, or train on your prompts, completions, or session data.
  • Require admin / org-owner credentials for personal use.
  • Auto-publish your numbers without an explicit opt-in toggle.
  • Lock features behind paid tiers in the MVP. Pricing comes after we have ≥1,000 active flexers.
  • Ship MCP tools that mutate state without surfacing the command they're about to run.
  • Estimate spend when we can measure it. Estimates only when the vendor leaves us no choice (and we say so).

Verify, don't trust

Every claim above is testable from your own terminal. Here's what to run if you want to audit the promises before we get them wrong.

# read the parser source — see exactly what we extract from your machine
npm pack tokenflexing@latest && tar -xzf tokenflexing-*.tgz && cat package/src/cli.js | less

# inspect what the daemon will write to disk (dry-run, no install)
npx tokenflexing@latest daemon

# preview the MCP wiring without applying it (dry-run, no edits)
npx tokenflexing@latest install-hooks

# see exactly which files the CLI reads on every scan
npx tokenflexing@latest scan

The CLI is the contract — the web app is just a viewer over the data you control.

Get on the board

Ready to flex?

One command pairs your machine + pushes the first sync. Takes about thirty seconds.