TokenEcon
Goal
Claude Pro meters you on two clocks at once: a rolling 5-hour session and a weekly limit, and other services (ChatGPT's Codex plans, OpenRouter and Groq free tiers) have allowances of their own. TokenEcon is a workbench for turning those meters into a pacing rule: how fast should automatic agents work right now so the week is spent well, not burned in two days and not left half unused when it resets? It is a mock-up and a model today, built so the rule can later be asked by real agents.
How it works
- On the Dial, type what Settings → Usage on claude.ai shows (the session and weekly percentages and their reset times). It answers burn, cruise, slow or pause, how hard agents may push, how much of this session and of the week they may use, and how many of the week's 33.6 five-hour slots are left.
- Overrides (burn faster, slow down, pause) and options (peak hours, overnight work, catch-up pressure) are switches on the same screen. The assumptions behind the numbers stay hidden until you ask for them.
- Week simulates a whole week three ways (no agents, agents with no governor, agents asking the model) with charts for the weekly curve, the 5-hour sessions, who used what, and what the model said each hour.
- Providers does the same job for other allowances, each switched on only when you track it. ChatGPT Plus and Pro (for Codex) go through the same pacing engine with their own settings: Plus has a 5-hour window and a weekly limit, Pro is reported to have only the weekly one. OpenRouter free models and the Groq free tier are counted differently: they are hard request and token limits per minute and per day, and anything unused is lost at the reset, so the rule is to spread the use out and spend what is left before the reset. A calculator shows what a given number of agents will do to a limit, and a ranking says where the next piece of work should go: first to whatever is about to expire unused, away from whatever is scarce.
- Agents holds concept agents you can switch on in that simulation, and the ask, work, report, sleep protocol they would follow.
- Models compares three versions of the decision rule on the same numbers.
- Lab is the reverse-engineering bench: a calibration log that estimates how much of the week one full session costs, an import for usage readings, and the theories being tested with what supports each.
- Roadmap lists the steps from this mock-up to agents that really ask.
Saving and privacy
Everything stays in this browser: your typed numbers, settings, calibration log, provider choices and agent choices. Nothing is sent anywhere and nothing reads your Claude account; you copy the numbers in yourself. Use Settings → Download backup to keep a copy. It is a working model, not a promise: Anthropic does not publish how the limits are counted, so every constant is a setting to calibrate.
Current state
Work in progress. The engine, the simulator and the provider calculations are finished and tested; the real numbers (how much of the week one session costs, the shape of your own week, each provider's exact limits and reset hour) wait for calibration. The limits for the other providers are starting values taken from public write-ups as of 2026-11-13, each with its source on the Providers tab, and providers change them without notice.
Changelog
- v0.2.1 (2026-11-13) — "My week" editor: under Show assumptions, choose "My own hours" and set how heavy each hour of your day is, plus a weekend factor, so the pace follows your real week.
- v0.2.0 (2026-11-13) — Providers tab: ChatGPT Plus and Pro (Codex), OpenRouter free and Groq free, with a quota calculator, routing across allowances, and the same agent contract and Telltale readings for all of them.
- v0.1.1 (2026-11-13) — renamed from Pacemaker to TokenEcon; the address is now /tokenecon (the old /pacemaker address redirects) and anything saved under the old name carries over.
- v0.1 (2026-11-13) — first workbench: engine with three model versions, week simulator with charts, concept agents, calibration Lab and roadmap.