Token Forecaster predicts how many output tokens a Claude Code reply will use, and its p90 line covered 90.6 percent of the calibration user's held-out calls

An agent turn that runs 5,000 tokens over what a person expected is worse than one that fails outright, because it uses context window, budget, and time before anyone can react. Sumcap Research launched Token Forecaster on Product Hunt on 25 September 2026, a free tool that predicts how many output tokens a Claude reply or agent turn is likely to use, before the request is sent and again while it runs.

The project is published at Sumcap/token-forecaster on GitHub under the MIT licence, and was built by Eduardo Nunes at Sumcap Research. It ships as a Claude Code status-line integration, a macOS menu bar app, a local dashboard, and a Chrome extension, with a "backoff ladder" that includes the model as one of the lookup dimensions. The tool watches requests and never modifies them.

What it actually predicts

Token Forecaster shows two numbers before generation starts: an ordinary length and a worst case, both in output tokens. While the reply is streaming, it labels the run as tracking within its ordinary band, running long, or running very long. The point, in the README's words, is to "reserve room in the context window, spot agent loops that are running long, and decide when a task is worth splitting". It does not force a length or set a token limit; it only reports.

The numbers, and the caveat

Sumcap publishes precise accuracy figures with the method attached. On a held-out set of 4,146 calls from one calibration user, "the p90 line covered 90.6 percent of held-out calls, 95 percent CI [89.7, 91.9]", and the p99 line covered "98.6 percent [98.2, 99.2]". Compared with a static reservation, the tool's forecast gave a "41 percent" improvement in pinball loss.

Sumcap names the critical caveat itself, and it is what makes the numbers usable. "One person's fit does not transfer." Holding the person fixed and swapping projects, the p90 coverage ran "95.4 / 80.9 / 88.1 / 88.2 percent" across four projects. Between users the calibration is unmeasured, and the company says the gap "is expected to be at least as wide as it is between one person's projects". A team adopting Token Forecaster should expect to go through the same calibration period the numbers above describe before the p90 becomes a reliable line to plan against.

Where it runs

Sumcap ships it for macOS 14 or later on Apple silicon and for Windows 10 1809 or later, both needing Node.js 22 or later. The Chrome extension loads unpacked. The status line is built specifically for Claude Code, with what Sumcap calls "experimental Codex support" alongside it.

The wider point

The measurement Token Forecaster shows a developer is one nobody in the LLM tooling stack was displaying directly. Anthropic's dashboards report cost and usage after the fact. IDEs show a running character count of the input. What has been missing is the piece Token Forecaster prints: a per-turn expectation of the output, calibrated to the person and the model, with a p90 and p99 line and a mid-run label.

A team running Claude Code past a few short sessions has a specific complaint the forecast answers. A subagent silently doubles the tokens the parent expected, an interactive turn slips into a long chain because the model kept rewriting a file, or a coding turn produces a 40,000-token diff that could not fit in the window. Each of those is visible in Token Forecaster before the reply is sent, and again while it is streaming, and the labels come with a p90 band that is honest about how well calibrated it is right now.

Source

Launch: Token Forecaster on Product Hunt.

Primary source: Sumcap/token-forecaster on GitHub.