We started with one question: can you tell how long an LLM reply will run before generation starts? Within a range, yes. On 4,146 calls the model had never seen, 90.6% came in under our worst-case line. Extended thinking stretched the tail 2.5x. Prompt wording was worth about 1%. The real driver is how long the agent loop runs, and nothing visible before the call knows that. An oracle for loop length would cut forecast error by 63%. Token Forecaster is what we built from those findings. Before you press Enter, it shows the usual length and a worst case. While the reply streams, it says typical, running long or very long. It runs in the terminal status line, a macOS menu bar app, a local dashboard and a Chrome extension. It only watches. It never changes the request, never sets max_tokens, and cannot save tokens or money on its own. It is for reserving context room, catching a loop that is running away, and deciding when a task is worth splitting. Scope today: calibrated on Claude Code history, with Codex ingestion experimental. The bundled profile comes from one developer's 16,687 calls. The daemon fits a personal profile from your own history and swaps it in only when it beats the bundled one. A second user reached 88.3% worst-case coverage on their own held-out calls. Everything runs locally: history is read, never modified, never uploaded. There is no installer yet. It builds from source on macOS (Apple silicon) and Windows. Every experiment that failed our gate sits in the README next to the ones that passed. Tell us where the forecast misses on your workload. Built by Eduardo Nunes at Sumcap Research.

Token Forecaster

Forecast how long an LLM reply will run, before Enter

Forecast how long an LLM reply will run, before Enter